v5.0.0

[5.0.0] — 2026-07-06

Plugin-native distribution + deterministic hook enforcement + Claude 5. v5 lands the full M1+M2+M3 scope of the skills-only 5.0 plan — this is the complete 5.0 release; there is no skills 5.1 (further MCP/memory work moves to the separate etyb.ai product).

Changed

• M1 (universal core): skills/etyb/SKILL.md rewritten to the Agent Skills open-spec subset: 967-char description (spec cap 1,024), 76-line body with the Tier 0-4 table inline — Tier 0 requests need zero additional file reads. core/signature.md, core/response-formats.md, core/scale-calibration.md, core/always-on-protocols.md consolidated into core/session.md; core modules total −40% words. Progress-marker and clarifying-question contradictions resolved (CTO-voice status lines, cap of 3, everywhere). Root AGENTS.md added (open standard); CLAUDE.md reduced to an @AGENTS.md import plus Claude-only notes. Whole-repo version 5.0.0 under the single-version policy.
• M2 (agents & hooks): five agent definitions under agents/ (explorer, planner, reviewer, stack-researcher, cartographer; memory: project) with model tiering — light work (explorer's search-and-cite, stack-researcher's fetch-and-distill) pins to the smallest model tier (model: haiku, mirroring the Codex emissions' gpt-5.4-mini); extensive-reasoning work (planner, reviewer, cartographer) inherits the user's session model. Pins only ever go down: the user-selected model is the ceiling, and any harness that can't resolve the field falls back to the session model. Hook scripts rewritten to the stdin-JSON contract (advisory systemMessage warnings, 21 tests) and wired via plugin hooks/hooks.json; deterministic adapter generator emits Codex TOMLs, Kiro agent JSONs, and a Cursor plugin skeleton under dist/adapters/. All five hooks verified firing against a real local Claude Code plugin install (see docs/plan/skills-5.0-plan.md M2-T4 for the observed transcript).
• M3 (remote stacks): manifest.json gains a generated 538-page stacks_pages array (scripts/build-manifest.sh); the stack-researcher agent owns the manifest-driven fetch + currency protocol; per-page currency checking with a batch-stamp detector (now hard-fails, CHECK_CURRENCY_STRICT=1); the anthropic-claude stack refreshed to the Claude 5 generation (Sonnet 5 / Opus 4.8 / Haiku 4.5 / Fable-Mythos tier) with per-page verification stamps, spot-checked against live Anthropic docs at release time. All 13 stack SKILL.md descriptions compressed under the 1,024-char open-spec cap (prose kept, trigger lists trimmed to the strongest signals); the token-budget lint now hard-fails over-cap descriptions.

Removed

• scripts/install.sh, scripts/update.sh, scripts/install-codex-runtime.sh and their tests — distribution is npx skills add e-t-y-b/etyb-skills (multi-harness) or the Claude Code plugin (.claude-plugin/plugin.json, name etyb, added).
• The three pre-release role skills (/etyb-review, /etyb-plan, /etyb-explore) and their Claude fork overlays, withdrawn before tagging. /etyb is the single trigger surface — the only skill a user (or auto-trigger) ever invokes. The same review/plan/explore work runs through the agents/ definitions, dispatched by the orchestrator into sub-agent contexts, so the user's session context stays clean. This also closed a review finding: the fork overlays only existed in dist/, so the installed role skills would have run inline — contradicting their own "fresh context" promise. Removing the surface removes the gap.

Fixed

• Docs no longer claim hook enforcement that is not wired — docs/installation.md's Enforcement Status note now accurately describes the advisory (non-blocking) hooks shipped on the Claude Code plugin path; the old lint check against gitignored .claude/settings.json — which could never pass on a clean clone — replaced with hook-script existence checks.
• New scripts/lint-token-budget.sh in CI enforces description caps; all 13 stack descriptions compressed under the spec cap (M3-T5).
• hooks/session-start-memory.sh shipped without the executable bit, which would have made the SessionStart hook fail with "permission denied" on a real plugin install; fixed.
• .claude-plugin/marketplace.json's plugin entry declared a redundant skills component list that conflicted with plugin.json's convention-based auto-discovery, causing claude plugin install to report the plugin as failed to load; removed.
• .gitignore's MARKETPLACE.md pattern (intended only for the repo-root internal doc) was unscoped and silently excluded stacks/vercel/marketplace.md from version control on case-insensitive filesystems; anchored to /MARKETPLACE.md and recovered the untracked page.
• Release-time truth sweep: CLAUDE.md, AGENTS.md, and README.md still said hooks and agent definitions "arrive in M2 — do not assume they exist" after M2 shipped them; all three now describe what actually installs. The marketplace plugin description double-counted the roster ("20 specialists + 9 protocols + 6 verticals" — the 20 already included the verticals); corrected to 14. A stale "Pro tier only" line in core/team-registry.md (tiers were removed in v4.0.0) dropped.