v1.4.2 — Claude Opus 4.8 adaptation + audit-driven hardening

Adapts the kit to Claude Opus 4.8 (released 2026-05-28) and lands an audit-driven hardening pass. The opus alias auto-routed to 4.8 with zero agent changes — this release catches up the human-readable surfaces and tunes agents/hooks to 4.8's behavior. Component counts unchanged (56 agents · 16 commands · 42 hooks · 22 skills · 19 rules).

🆕 Opus 4.8 adaptation

• Version refresh 4.7 → 4.8 (badge, CLAUDE.md templates, docs, showcase trailers).
• Effort convention documented (per-role xhigh/max; default high) + escalation to max on a failed /dev gate/retry + a fast inline self-review before the expensive subagent gates.
• Reviewers shifted to coverage-not-filtering — a two-stage discovery→triage discipline with an "Open Questions" bucket, so 4.8's more literal instruction-following no longer drops real low-severity findings.
• Added a shared Evidence Gate (cite-line / concrete-trigger / context-checked / defensible-severity) and dropped the legacy "think step by step" idiom.
• Verifier discipline ("no approval without fresh evidence", hedge-word auto-reject) + a standardized 3-state verdict (PASS / NEEDS-ATTENTION / NEEDS-REMEDIATION) in goal-verifier and gan-evaluator.

🐛 Verified bug fixes

• [wip]-on-main protection was dead code ($GIT_CMD_PREFIX used before definition) — fixed + regression test.
• Four discipline hooks keyed state on bare $PPID and silently never accumulated — migrated to a shared superkit_session_key() (proven by a cross-call test).
• audit-trail hash chain was a single global file — now per-session so parallel subagents don't cross-link it.
• superkit-counts-verify convenience wrapper hung / silently no-op'd in non-TTY contexts — now always runs the real check.
• Reviewer audit-mode reframed off unexecutable "self-spawn" wording (subagents can't spawn sub-agents on Claude Code 2.1.x) + removed the dangling Agent tool from leaf reviewers.

🛠️ Hooks & statusline

• Warn-only PreToolUse advisories now reach the model via hookSpecificOutput.additionalContext.
• Per-session advisory dedup/throttle (CLAUDE_NUDGE_WINDOW_MIN) — 4.8 self-flags, so lighter nudging.
• superkit-counts-verify hardened (label-anchored parsing, single-source internal-file exclusions, network check gated behind --check-remote).
• statusline additively surfaces effort level + context budget when the CLI provides them.

🤝 Codex + showcase sync

• Every reviewer modernization (two-stage discipline, Evidence Gate, idiom removal, 3-state verdict, audit-mode reframe) mirrored into the Codex skills and showcase — keeping all surfaces consistent.

📝 README & docs

• README reworked: one-line value wedge, unified install flow, dual Claude-Code+Codex pillar, a "which command do I use?" decision table, a concrete /dev walkthrough, and an all-Opus rationale.
• Swept stale counts/config across the guides, INSTALL, and Codex docs (Codex skills 72→82, core hooks/skills, /dev phase counts, config.toml gpt-5.5/xhigh, Stop-hook model→opus, showcase rules 3→6).

✅ Quality

• 11 hook test suites + the full node/install smoke suite green; counts/version verified; two full all-Opus multi-agent audits (functionality + documentation) drove the fixes above.

Known follow-up: the 5 hand-drawn /dev flow SVGs still render 15 nodes (canonical is 16) — a pre-existing diagram re-authoring task, tracked separately.

🤖 Generated with Claude Code