AlexZio00/sovereign-skills

README preview

🌐 **English** | [한국어](docs/README.ko.md) | [日本語](docs/README.ja.md) | [中文](docs/README.zh.md) | [Español](docs/README.es.md)

# sovereign-skills v6.5.7

20 skills for the full Claude Code project lifecycle — from setup to daily workflow to code review to session management to governance. Each skill is useful standalone; the full sequence covers everything.

> **What changed in v6.5.7:** Refinement release — no skills added or removed; several gained working deterministic scripts in place of LLM self-scoring. `project-overview` (`generate_overview.py` is no longer a stub — fully implemented registry parsing, state-snapshot extraction, and AUTO-block render/replace, with untrusted-input table-cell escaping and malformed-marker recovery, backed by unit + integration tests), `scope` (Quick/Full ambiguity gating and BRIEF.md min-item validation now run through a ported `ambiguity_gate.py`, with a regression test locking a bold-text-as-header counting bug), `skill-ops` → v1.2 (Health Mode bucket classification and Quality Mode S/U/S_Q scoring now run through `skill_health_bucket.py` instead of manual arithmetic), `collab-audit` (Step 0.6 source-hygiene filtering now runs through `session_hygiene_scan.py`), `session-checkpoint` (new Discoverability Check step flagging memory writes with no index backlink, a timeout/kill partial-output guard, two Reflexion lesson-quality gates, a PII-redaction rule for raw observations, and Key Files verification moved to a `validate_memory_claims.py` script), `goal-lock` (scope-check surface widened to interface/API changes not just touched files, a benchmark-backed chain-length dominant-variable note, the Stop-hook order gate upgraded to a verified 4-condition implementation, a new Safety Layers section, enum-style VERIFY failure labels, a self-judgment caveat on REFINE's DELTA CHECK, and a background-task termination handshake), `code-autopsy` (a Rationalization Table of 8 common reviewer rationalizations + rebuttals, and a note recommending script-based severity arithmetic over mental math), `eval-leakage-audit` (17→18-pattern taxonomy — adds Goodhart co-evolution in self-improving loops — plus a Reviewer Independence Honest 4-Label check), `integration-intake` (read-only `tools:` frontmatter, redirect-style `not_for` entries, 7 fixed misjudgment-enum labels per phase), `pre-push` (explicit `depends_on`/`concurrency_profile` frontmatter, an Autonomy Boundary note distinguishing read-only git commands from the gated `git push`, an external security-catalog reference link), `session-start` (`depends_on`/`concurrency_profile` frontmatter, Phase 2.2-2.4 rewritten to deterministic commands with fixed stdout contracts, a new unnumbered L0 Inheritance section), `doc-drift` (new Step 0 deterministic pre-filter feeding the Risky/Ambiguous category as supporting evidence only), and `depends_on`/`concurrency_profile` frontmatter adopted across `full-audit`, `project-check`, `project-init`, `freeze`, `stepback`, `next-action`, and `clean-room`, several of which also gained tool-category-tagged Scope Boundary tables. **Known gap surfaced by this release**: `project-check` (and a few siblings) gained unnumbered "inherited principle" notes on Safety Layers/Error Recovery headers — a past release deliberately stripped numbered `(L0 §N)` citations from these same headers across all 10 skills; this version's notes are unnumbered/generic rather than the removed citations, but the header-annotation pattern itself is back. Worth a conscious call before the next release builds further on it.
>
> **What changed in v6.5.6:** Refinement release — no skills added or removed. `eval-leakage-audit` (13→17-pattern taxonomy — respawn masking, pseudo-replication, stimulus-calibration gap, unaudited cost-saving skips), `goal-lock` (an S7 stop rule that blocks forcing an implementation through when execution evidence contradicts an explicit instruction, a 5-tier evidence-rigor ladder + failure-first reporting order + banned hedge phrases, a "layer laundering" success-masquerade pattern, and an evidence-rigor pre-spec note), `full-audit` (a composite-accumulation-gate that flags death-by-thousand-cuts risk when UNCERTAIN/NIT findings pile up in one area even though each was individually dismissed, and an Assumption Ledger for CONFIRMED verdicts that rest on unverified assumptions), `session-checkpoint` (optional `regime`/`escalate_if` lesson fields, an optional `outcomes` field tracking first-attempt-pass/rework/resolvedBy, and a second lessons-archival OR-condition plus a mandatory cross-reference check before archiving), `code-autopsy` → v7.2 (a code-smell vocabulary for Q1, a concrete-failure-scenario requirement, object-level authorization under Q5, expanded network/DB/streaming/cache checks under Q7, a shallow-module + ADR-conflict check under Q8, a numeric confidence-threshold system, and an outcome-ceiling-to-process-metric switch for tied comparisons), `pre-push` → v3.8.0 (ships `scan_secrets.py` alongside the existing `scan_secrets.pl` — Python preferred when present, Perl as fallback — plus `LICENSE.txt` for upstream attribution; note: the two scanners' pattern coverage was found to differ in this release, e.g. the Python port adds an f13 Slack-webhook check absent from Perl, and several existing patterns cover a narrower or wider set of variants on each side — documented as a known gap, not yet reconciled).
>
> **What changed in v6.5.5:** Refinement release — no skills added or removed. `eval-leakage-audit` (8→13-pattern taxonomy — dual-fail-flag, asymmetric-baseline self-falsification, evidence-burn, an ungraded-grader 4-gate check, ceiling-task detection — plus a stratification-substantiation checklist and an honest reviewer-independence 4-label verdict), `goal-lock` (an order gate that blocks completion when no verification ran after the last edit, evidence-channel branching for non-exit-code deliverables, a comprehension check, and three new success-masquerade guards), `pre-push` → v3.7 (cross-bundle joint pass + a three-state false-positive gate), `integration-intake` (headroom pre-classification, trait-vs-procedure termination, triple-check effectiveness claims), `session-start` (added `claude-sonnet-5` to the model-ID allowlist — fixed a false "invalid model" warning), `setup` (removed a duplicate `/setup` trigger), `full-audit` (a rule-dry-run verification layer), `clean-room` (reconciled with upstream autobahn v0.14.0 — an N=1 cap on the independent re-sweep), `code-autopsy` (Q10 oracle-redefinition detection), plus smaller refinements to `session-checkpoint`, `skill-ops`, `collab-audit`, and `scope`.
>
> **What changed in v6.5:** New: `eval-leakage-audit` (audits whether an eval/metric/holdout actually secures independent external ground truth vs circular self-confirmation, via an 8-pattern taxonomy; read-only), `doc-drift` (audits the memory/docs Claude Code loads into context — CLAUDE.md/MEMORY.md/skills/agents/commands — for outdated claims, mutual contradictions, and risky/ambiguous wording, producing a prioritized fix list). Updated: `project-init` (fixed a `skill.md`→`SKILL.md` filename-casing bug that could break skill loading on case-sensitive filesystems, and moved the Phase 3 templates into `references/templates.md`), `pre-push` → v3.6 (two new secret-scanner patterns — f11 prompt-injection strings in diffs, f12 non-PyPI supply-chain index URLs — plus a Step 0 Hook Pipeline Health check), `scope` (added the Mid-Task Scope Drift 10x-Discovery Rule), `collab-audit` (added Step 0.6 Source Hygiene Filter to exclude auto-derived subagent/thread sessions), `full-audit` and `integration-intake` (both gained a Safety Layers section; `integration-intake` also added a Phase 1.8 M-axis surface-selection step — judge which surface (prompt/rule/hook/skill) a pattern should live on before routing), `goal-lock` (added a `migration` task template), `project-overview` (added a Rationalization Table), `stepback` (added a Dominant Variable section + frontmatter fields).
>
> **What changed in v6.4:** New: `full-audit` (exhaustive area audit — deterministic sweep + content review, persistent coverage map, anti-false-positive kill-test), `integration-intake` (5-item screening gate for adopting external skills/agents/rules/plugins, with a provenance/injection check), `clean-room` (carves safety-adjacent requests into a safe scope executed by a genuinely isolated fresh-context subagent, adapted from LilMGenius/paperthin's "autobahn" skill under MIT with a filesystem-isolation and ledger-timing upgrade). Updated: `goal-lock` (a constraint re-echo check at long-task checkpoints, so CONSTRAINTS/SCOPE-Exclude don't quietly fall out of view during extended work), `session-checkpoint` (a new Attestation phase — an evidence-chain receipt log with a bundled `handoff_attestation.py`, so the next session's SessionStart hook can detect handoff tampering).
>
> **What changed in v6.3:** New: `skill-ops` (snapshot/rollback + usage health + invocation tracking hub), `next-action` (reads handoff/git/lessons/STATE and proposes the top-3 next actions), `project-overview` (deterministic cross-project status map). `code-autopsy` → v7.1 (deeper sub-checks per question), `pre-push` → v3.5 (9 supply-chain IOC patterns), `goal-lock`/`session-checkpoint`/`session-start`/`scope`/`stepback`/`freeze` all strengthened. All 12 prior skills gained `not_for` and `see_also` frontmatter for better discoverability.

---

View full repository on GitHub →