meta-skills v4.0.0 — Enterprise Quality Engine for Claude Code

Enterprise-grade plugin for Claude Code: 16 skills, 17 commands, 9 hooks, 6 agents, 27 scripts. Implements all 7 research principles. Adversarial review. CI/CD gates. Cross-model refinement.

Install

# Register as local marketplace (once)
claude plugins marketplace add ./meta-skills --scope local

# Install plugin (survives restarts)
claude plugins install meta-skills@meta-skills-local --scope local

# Verify
claude plugins list | grep meta-skills

Architecture

meta-skills/
  .claude-plugin/plugin.json       # Plugin manifest (v4.0.0)
  hooks/
    hooks.json                     # 7 events, 12 hooks
    lib/config.py                  # Centralized settings (all tunable values)
    lib/services.py                # Shared clients (Honcho, open-notebook, vault)
    lib/hook_wrapper.py            # Shared hook utilities
    session-init.py                # First-prompt: Honcho + open-notebook + CI + watcher
    session-stop.py                # Auto-summary + Honcho + KB recommendation + P7 state
    correction-detect.py           # Correction detection + S10 compliance
    scope-tracker.py               # Topic drift advisory (3+ switches)
    approach-guard.py              # Unauthorized strategy switch blocker
    exploration-first.py           # Read-before-write enforcer (P5)
    quality-gate.py                # Test/lint failure gate + commit/push checks
    token-audit.py                 # Per-tool-call token measurement (JSONL)
    meta-loop-stop.py              # Objective iteration loop gates
  skills/                          # 16 skills with SKILL.md + references/
  commands/                        # 17 slash commands
  agents/                          # 6 sub-agents (doc-auditor, doc-editor, 3x scanner, session-analyst)
  scripts/                         # 27 deterministic Python scripts
  self-improving/                  # Plugin-local learning (memory.md, corrections.md)
  oversight/                       # Quality snapshots, calibration, autoreason results
  plans/                           # Implementation plans

Skills (16)

Skill	Purpose
creator	Cooperative skill creation (5 phases)
design	Visual DESIGN.md generator
dispatch	Intelligent skill routing
doc-updater	Documentation sync orchestrator
feedback	Bidirectional end-of-session review
git-worktrees	Parallel branch workflow
harden	Automated SCAN-TRIAGE-FIX-VERIFY-REPORT loop
init	Project entry point (audit/goal/setup)
judgment-day	2 blind judges, convergence pattern
knowledge	Knowledge funnel (log/search/sync/audit)
refactor-loop	Scan-Improve-Verify cycle (one change per iteration)
statusbar	Session lifecycle (statusline + watcher + sync)
systematic-debugging	Root-cause analysis framework
tdd	Test-driven development workflow
triad-review	3-perspective adversarial review
verify	No completion without evidence (Iron Law)

Commands (17)

Command	Purpose
/meta-create	Cooperative skill creation
/meta-design	Visual DESIGN.md generator
/meta-discover	Session pattern analysis → skill suggestions
/meta-docs	Doc sync via agent team (presets: quick, infra, full)
/meta-feedback	End-of-session review
/meta-knowledge	Knowledge funnel operations
/meta-audit	Skill audit (usage, staleness, efficiency)
/meta-harden	Automated hardening scan + fix
/meta-judgment	Adversarial judgment day
/meta-ci	CI/CD status dashboard
/meta-loop	Objective iteration loop with real gates
/cancel-meta-loop	Stop active meta-loop
/meta-quality	Quality snapshot
/meta-snapshot	Full plugin state snapshot
/meta-status	Plugin health check
/meta-test	Behavioral skill testing
/meta-triad	3-perspective adversarial review

Hooks (9)

Hook	Event	Addresses
session-start	SessionStart	Honcho, open-notebook, CI check, watcher spawn
session-init	UserPromptSubmit	Prompt counter + P7 context recovery
correction-detect	UserPromptSubmit	Correction patterns + S10 compliance
scope-tracker	UserPromptSubmit	Multi-task drift advisory
approach-guard	PreToolUse (Bash)	Wrong Approach blocker
exploration-first	PreToolUse (Write|Edit)	Read-before-write + write-time QA (P5)
token-audit	PostToolUse (all)	JSONL logging per tool call
quality-gate	PostToolUse (Bash)	Test/lint failures + commit gate + push CI
context-recovery	PreCompact	State snapshot before context compaction
meta-loop-stop	Stop	Objective loop gates
session-stop	Stop	User-facing verification + guidance
session-end	SessionEnd	Honcho write + state persist + cleanup

7 Research Principles

#	Principle	Implementation
P1	Confidence-Weighted Consensus	Borda count with high/medium/low confidence → verdict levels
P2	Behavioral Tests	test-scenario.md per skill, pass/fail regex validation
P3	Orthogonal Revision	Author C generates fundamentally different approach; recombine B+C→D
P4	Correction Promotion	User corrections → persistent rules (corrections.md → CLAUDE.md)
P5	Write-Time QA	exploration-first hook: 3 reads before first write
P6	Cost Routing	Model assignment per task complexity (haiku→sonnet→opus)
P7	Context Recovery	Prompt counter + state sentinel, survives compaction

Quality System

Component	What	Inspired by
harden	Automated SCAN-TRIAGE-FIX-VERIFY-REPORT loop	sd0x-dev-flow, Citadel
judgment-day	2 blind judges parallel, convergence pattern	gentle-ai
quality-gate	Auto-detect test/lint failures + commit gate	Plankton, pilot-shell
meta-loop	Objective iteration loop with real gates	ralph-loop
refactor-loop	Scan-Improve-Verify cycle (one change per iteration)	adversarial-dev
verify	No completion without evidence (Iron Law)	superpowers
autoreason	Cross-model refinement (7 CLIs, Confidence Borda, Orthogonal Revision)	NousResearch/autoreason
behavioral-tests	test-scenario.md per skill, pass/fail regex validation	OpenJudge Skill Graders
context-recovery	Prompt counter + state sentinel, survives compaction	sd0x-dev-flow

Scripts (27)

Core Quality

Script	Purpose
harden.py	Frontmatter validation + automated fix
autoreason-skills.py	Cross-model adversarial refinement (7 CLIs)
test-skill.py	Behavioral test runner
eval.py / eval-skill.py	Quality scoring (0-100)
validate.py	CI gate (frontmatter validation)
ci-status.py	CI/CD status monitor

Session Lifecycle

Script	Purpose
statusline.py	Rainbow statusbar (model, cost, context)
session-watcher.py	Per-session guardian (RAM warning, ghost cleanup)
process-monitor.py	System-wide process monitor
benchmark-session.py	Token benchmark (before/after comparison)
token-report.py	Token efficiency analysis from audit data

Plugin Management

Script	Purpose
plugin-setup.py	First-run setup (auto/interactive, cross-platform)
build-skill-registry.py	Auto-generate skill registry
project-scan.py	Project scanner (stack, files, LOC, quality)
quality-snapshot.py	Full plugin quality snapshot
oversight.py	Oversight report generator
migrate-frontmatter.py	Frontmatter migration tool
promote-corrections.py	Promote corrections to rules
reworker.py	Auto-fixer (diagnose + fix score problems)
filter-meta.py	Metadata filter utility
setup-meta-loop.py	Meta-loop setup
session-end-sync.py	End-of-session sync helper

Configuration

Centralized settings in hooks/lib/config.py. Override via ~/.claude/plugins/data/meta-skills/config.json:

{
  "features": {
    "watcher": true,
    "correction_detect": true,
    "scope_tracker": true,
    "approach_guard": true,
    "exploration_first": true
  },
  "thresholds": {
    "min_reads_before_write": 3,
    "consecutive_failures_warn": 3,
    "scope_drift_warn_switches": 3,
    "correction_pause_count": 2,
    "context_recovery_gap": 10
  },
  "autoreason": {
    "num_judges": 3,
    "max_passes": 5,
    "convergence_k": 2,
    "cli_timeout_s": 180
  },
  "quality_gate": {
    "block_commit_on_lint_fail": false,
    "block_push_on_ci_fail": false,
    "warn_commit_without_lint": true
  }
}

CI/CD

GitHub Actions workflow (plugins-ci.yml) with 5 gates:

1. Syntax — py_compile all Python files
2. JSON — plugin.json validation
3. Hook Safety — exit 0 + crash-safety checks
4. Skill Validation — frontmatter + body length
5. Harden Scan — harden.py --scan (0 CRITICAL required)

Self-Improving System

• self-improving/memory.md — Preferences, patterns, learned rules
• self-improving/corrections.md — Mistakes not to repeat (promotes to rules)
• oversight/ — Quality snapshots, calibration data, autoreason results

Services Integration

Service	Purpose
Honcho	Cross-session user context (peer detection, derived summaries)
open-notebook	Knowledge base (RAG search, source creation)
GitHub Actions	CI/CD gates (5 workflows)

v4.0 Changelog

• Phase 1: Centralized settings (hooks/lib/config.py)
• Phase 2: English translation (all hook output messages)
• Phase 3: Centralized state manager (hooks/lib/state.py — replaces 7 patterns)
• Phase 4: Hook event expansion (4→7 events, 9→12 hooks)
• Phase 5: Command standardization (all 17 → Pattern A or B, <10 lines each)
• Phase 6: Skill frontmatter schema (type/category/requires on all 16)
• Phase 7: Documentation + v4.0.0 release

License

MIT

Author

AI Engineering — kontakt@ai-engineering.at