v6.2 — code-autopsy + stepback + 10 skill upgrades + Codex support

What's New in v6.2

New Skills

• code-autopsy — 12Q quantified code review. 4-axis scoring (Security × 0.35 + Stability × 0.30 + Robustness × 0.20 + Operability × 0.15). Severity Anchor Table. Deployment verdict (SHIP/FIX/RISKY/BLOCK). CapCode ceiling metric for score gaming detection. CEF fabrication detection. Quick mode and Diff mode. Works as a standalone prompt in any LLM.
• stepback — One-shot perspective reset. 1 abstract reframing question + 3 quick checks (scope drift, side effects, better approach) in under 10 lines. Use anytime during implementation.

All 10 Existing Skills Upgraded

• goal-lock — Ralph Wiggum early-completion detection (12th masquerading pattern) + verification traceability (every claim must trace to an actual tool call)
• collab-audit — 13→14 sections. New: Thinking Level Trajectory (5-Level model + AI attribution correction)
• session-checkpoint — Handoff clarity self-check (2 anchor questions after writing)
• session-start — Context rot prevention (sliding window for stale entries)
• pre-push — 3-IOC supply chain check for newly added dependencies
• scope — Contraindication field (conditions where the chosen approach is NOT suitable)
• freeze — Thaw protocol (formal unfreeze workflow with blast radius check)
• project-init — Extended .env.example template (OAuth, external services, monitoring) + security baseline
• project-check — Score delta tracking (compare current vs previous scan results)
• setup — Redesign protocol for Tier 0 violation test failures (3-option escalation)

Infrastructure

• Repo renamed: claude-code-skills → sovereign-skills
• Codex/Cursor support: agents/openai.yaml for all 12 skills. Install via npx skills add AlexZio00/sovereign-skills --skill <name> --agent codex
• README in 5 languages: English, Korean, Japanese, Chinese, Spanish
• Internal references removed: All provenance/source attributions moved to central index. Clean public release.

Lifecycle Coverage (12 skills)

Setup:     /project-init → /setup
Workflow:  /scope → /freeze → /goal-lock → /stepback → /code-autopsy → /pre-push
Session:   /session-start ↔ /session-checkpoint
Quality:   /project-check, /collab-audit

Install

git clone https://github.com/AlexZio00/sovereign-skills.git
cd sovereign-skills
for d in */; do [ -f "$d/SKILL.md" ] && cp -r "$d" ~/.claude/skills/; done

MIT License. DM @AlexZio00 for custom skill development.

v6.0 — Consolidated & Upgraded

What's New

Added

• goal-lock — Agent discipline engine. Locks the goal, enforces PLAN→DO→VERIFY→FINALIZE→OUTPUT loop, detects 11 success masquerading patterns (test deletion, mock wrapping, threshold relaxation, etc.). Quick mode (3 fields) for small changes, Full mode (7 fields) for everything else.

Merged (13 → 10 skills)

• harness-init + team-init → setup — Infrastructure and agent team in one guided flow
• brief + adr → scope — Scope definition with ADR capability built in
• retro → session-checkpoint — Retrospection is now Phase 1.7 Reflexion

Removed

• token-audit — Use npx ccusage directly
• adr (standalone) — Absorbed into scope
• retro (standalone) — Absorbed into session-checkpoint

Upgraded (all 10 skills)

• Dominant Variable, Key Assumptions, Error Recovery, Safety Layers added to every skill
• Scope Boundary with action tags ([READ]/[WRITE]/[BASH]/[AGENT])
• session-checkpoint: Memento CoT compression, Reflexion, Invocation logging
• pre-push: Large diff deterministic bundling, Discard If conditions
• collab-audit: Anti-pattern flags, Key Assumptions

Install

git clone https://github.com/AlexZio00/claude-code-skills.git
cd claude-code-skills
for d in */; do [ -f "$d/SKILL.md" ] && cp -r "$d" ~/.claude/skills/; done

v5.0 — Three new skills + session health check

New Skills (3)

/freeze — Scope lock before implementation

• Parses your input to classify files as editable, frozen, or read-only
• Emits a FROZEN SCOPE block and stops — no code generation, no agents
• Ambiguous scope → one clarifying question; still vague → freeze the broader scope
• Adapted from garrytan/gstack freeze pattern (task-scoped, no session state)

/retro — Milestone retrospective

• 6-phase pipeline: scope → friction points (root cause) → wins → pattern extract → lessons write → summary
• Duplicate check before writing — updates obs/seen on existing lessons.md matches
• Complements session-checkpoint Reflexion (session-scoped); retro is milestone-scoped and goes deeper

/token-audit — Token overhead measurement

• Measures actual overhead from your session JSONL — not estimates
• Turn-weighted averaging across 8 categories (config baseline, memory re-reads, skill loading, agent overhead, etc.)
• Optional dark-theme infographic with 6 panels (pip install matplotlib)
• Benchmark: standard setups 35–50% overhead, heavy agent systems 50–65%

---

Enhancements

/session-start — Phase 0.5: Environment Health Check (runs every session)

• Checks ~/.claude/settings.json model ID against known-valid Claude identifiers
• Counts permissions.allow entries — warns if > 5 (silent permission surface growth)

/project-check — Safety Layers + Error Recovery

• Explicit table of blocked actions mapped to Invariants — read-only guarantee is now auditable
• Failure classification table: partial scans report as partial, never as complete

Usability fixes across the suite:

• /pre-push: perl prerequisites documented (Strawberry Perl for Windows)
• /session-checkpoint, /retro: auto-create tasks/lessons.md with header if missing
• /harness-init: mkdir -p tasks/ guard before lessons.md write
• /brief: trigger collision with /freeze resolved ("scope lock" moved to freeze)

---

Installation

SKILLS_DIR=~/.claude/skills
mkdir -p $SKILLS_DIR/{brief,adr,project-check,project-init,harness-init,team-init,pre-push/scripts,collab-audit,session-start,session-checkpoint,freeze,retro,token-audit}
# copy each skill directory's SKILL.md — see README for full commands

Full install commands: README → Installation

---

13 skills total. MIT.

v4.7 — Pre-push resilience + session improvements

What's new

pre-push v3.2.0

• Agent failure recovery — if a review agent times out or errors, retry once. Still failing → report ⚠️ SKIPPED (agent unavailable — {agent}) in the summary and continue. Never silently treat a failed agent as PASS.
• Conflict resolution — when agents give opposing verdicts on the same file, the strictest wins (weakest-link principle). Fully opposing verdicts (PASS vs Critical FAIL) → report both to user, do not push.

session-checkpoint

• Snapshot cleanup — detects entries older than 90 days in ~/.claude/.harness/snapshots/ and notifies the user. Not auto-deleted — manual review required.
• STATE.md integration — new Phase 3 step: update cross-project state file when blockers are resolved or milestones land. Phase 4 checklist gains a STATE.md review item.

session-start

• Phase 4.1 Selective Load — tag-based MEMORY.md filtering. <!-- #always --> sections load in full; <!-- #on-demand --> sections surface as TOC only (Grep when needed). No tags → full load (backward compatible). Reduces context overhead as MEMORY.md grows.

Cleanup

• Removed internal framework labels from Safety Layers and Truthful Reporting section headers across all 10 skills. No behavior changes.

Update

git pull
cp pre-push/SKILL.md ~/.claude/skills/pre-push/SKILL.md
cp session-checkpoint/SKILL.md ~/.claude/skills/session-checkpoint/SKILL.md
cp session-start/SKILL.md ~/.claude/skills/session-start/SKILL.md

v4.0 — 10 Skills, 4 Categories

What's New in v4.0

3 New Skills

• session-start — load handoff, review lessons, ready signal at session open
• session-checkpoint — extract entities, rewrite handoff, save memory at session close
• adr — capture architecture decisions while building (pairs with /brief)

Renamed

• -> (naming consistency with session-start)

README Restructured

• 4 categories: Project Setup / Daily Workflow / Session Management / Quality
• 10-skill lifecycle table with Phase/Frequency/Benefit
• Quick Start section with daily session loop
• 4-layer memory system explained

Install

bash:

Windows:

If upgrading from v3.x: is now . Update your trigger keywords accordingly.