v1.1.0 — README redesign & quick-start visuals

v1.1.0 — README redesign & quick-start visuals

A documentation-only release. No code, schemas, or behavior changed since
v1.0.0 — the Meta-Skill regression suite still reports 121 passed. This
release polishes the project's front door: a fully restructured bilingual README
and three custom-generated quick-start diagrams.

SemVer: minor bump. The change set is documentation and assets only, with no
breaking changes and no new code surface — a minor version signals a visible,
user-facing improvement to the project's presentation.

What changed

README — full restructure (EN + 简体中文, 1:1 mirror)

The previous README had grown by accretion to 22 sections / 638 lines, and the
reliability section at the end read as a labeled patch rather than an integrated
part of the project. Both READMEs were rewritten from scratch:

• Leaner narrative — 638 → 399 lines (EN), 593 → 400 lines (ZH), modeled on
  the structure of trending open-source READMEs (hero → why → quick-start-up-front
  → how-it-works → features).
• Human usage front-loaded — a 3-step guide (Prepare → Run → Review) now sits
  right at the top of Quick Start, with an ASCII-free visual and a plain-language
  "what happens behind the scenes" explanation.
• Reliability, integrated natively — the multi-judge review, dry-run gate,
  cross-process crash recovery, and tamper-evident audit are now woven into the
  How It Works section as built-in guarantees, with no internal jargon
  (T3 / T4 / WP-13) leaking into the reader-facing flow.
• Deep dives collapsed — repository layout and key-docs tables moved into
  collapsible <details> blocks; the slide-walkthrough meta-section was removed
  in favor of a link to the deck.

Quick Start — three new diagrams (Grok-generated)

The three ASCII-art boxes in Quick Start were replaced with real flat-vector
infographics, generated via the Grok CLI and color-matched to the existing
22 sanitized deck images (emerald #10B981 / #059669 + slate #6B7280 +
white) so the new art does not clash with the established visual identity.

Image	Replaces	Shows
qs1-3step-flow.png	3-step ASCII logic diagram	PREPARE → RUN → REVIEW (middle highlighted)
qs2-prepare-task.png	(new, above the directory tree)	SKILL + USAGE SET + CONFIG → TASK
qs3-behind-scenes.png	linear behind-the-scenes arrow diagram	RUNNER → EVAL → DIAG → OPTIMIZE closed loop, with a blind-wall/lock isolation visual

Design choices:

• Language-neutral — art carries only short English labels; long-form
  explanation and the copyable directory tree stay in markdown/code blocks.
  One image set serves both locales, so there is no bilingual image maintenance.
• Reproducible — the exact Grok prompts and the locked design spec are saved
  in docs/images/prompts/quickstart-style-spec.md; SVG sources for two of the
  three diagrams are kept alongside the PNGs for future tweaks.

Verification

Meta-Skill native regression
  121 passed

working tree
  clean, in sync with origin/main

code diff since v1.0.0
  0 .py files changed

Compatibility

Fully backward-compatible. No breaking changes. Nothing that consumed v1.0.0
needs to change for v1.1.0.

---

Full diff: v1.0.0...v1.1.0

v1.0.0 — First release: Meta-Skill + SES, WP-13, final remediation

v1.0.0 — First release

The first tagged release of Skill Optimizer · Meta-Skill with the
Skill Evolution System (SES) system layer. This bundles the full
Champion/Challenger optimization pipeline, the WP-13 Darwin/SkillLens/SkillOpt
alignment work, and the Round 12 final remediation that closed every
previously-deferred reliability gap.

What this release is

A Meta-Skill (a skill that optimizes other skills) that drives an
auditable, fail-closed, crash-recoverable optimization loop against a real
Human Reference Usage Set:

test → evaluate → diagnose → patch → retest → regression → validation → certified release

Headline features

Meta-Skill (the optimization protocol)

• 🔁 Closed loop with 3-layer evaluation: Hard Assertions (veto) → Quality
  Rubric (scored) → Pairwise Blind (head-to-head).
• 🛡️ Anti-overfitting: hardcoding detection, 70/30 Optimization/Validation
  split, gold-free diagnosis.
• 🌐 Cross-agent: open SKILL.md + standard-library Python; binds to no SDK/MCP.
• Regression suite: 121 passed, 70 subtests (standard library, offline).

Skill Evolution System (the system layer) — ses package

• Explicit, fail-closed state machine (26 states, 105 transitions) with a
  tamper-evident hash-chained audit log and content-addressed evidence.
• Certifying sandbox isolation, independent Hidden-Holdout trust domain,
  and a GitHub-SHA-pinned promotion gate.
• Real cross-process crash recovery — a crashed process resumes from a
  verified checkpoint without rerunning the baseline, recreating candidates,
  or duplicating bridge steps.

WP-13 alignment (Darwin / SkillLens / SkillOpt)

• T1 deterministic skill-quality checks (failure-mechanism, actionable-
  specificity, high-risk-action blacklist).
• T2 configurable, auditable early stop ("stop when gains vanish").
• T3 independent variant-review judge pool — ≥2 certified Reviewers per
  round with conservative aggregation and cross-round session isolation.
• T4 dry-run reliability gate — blocks a round whose Runner dry-run
  ratio exceeds the frozen threshold, before the Holdout is consumed.
• T5 destructive-command / rollback-safety guard.
• T6 SkillOpt alignment ADR (research-only; do-not-adopt conclusion).

Round 12 final remediation (all gaps closed)

• Global-best candidate selection (KEEP/REVERT per round; finalization restores
  the selected attempt, not the final attempt).
• finalize_failed terminal semantics; verified checkpoint-hash recovery;
  real cross-process recovery; accurate per-attempt candidate status.

Verification

Check	Result
Full SES suite	847 passed, 5 skipped
Meta-Skill native regression	121 passed, 70 subtests
ruff / mypy / compileall	All passed
scripts/acceptance.sh	ALL OFFLINE ACCEPTANCE CHECKS PASSED
git whitespace (diff --check)	clean
External production certification	NOT EXECUTED (no Docker / GitHub Enterprise / production-model credentials)

Full evidence: docs/skill-evolution-system/wp-13-final-remediation-acceptance-report.md.

Getting started

git clone git@github.com:nxl801/skill-optimizer.git
cd skill-optimizer
python3 -m venv .venv-ses && .venv-ses/bin/pip install 'pydantic>=2' fastapi 'uvicorn[standard]' httpx pytest ruff mypy
make ci          # tests + lint + preflight
make demo        # offline E2E demo

The skill lives at .agents/skills/optimizing-skills-from-human-usage/.
Review reports/final-report.md and ship only optimized-skill/.

Known limitations

• External certification is NOT EXECUTED. Docker image build/digest
  verification, GitHub Enterprise protected-branch workflow, and production
  RuntimeBroker dispatch are environmental steps not run in this release.
• Output is text-first. The contract evaluates one UTF-8 text deliverable
  per execution; multi-file/binary artifacts need a deterministic textual
  comparison artifact.
• Host isolation is external to the skill format; this package specifies
  and validates the protocol. A host that cannot enforce role boundaries must
  emit SAFETY_BLOCKED, not a lower-integrity single-agent run.

---

Tag: v1.0.0 · Commit: 777e5a0 · Branch: main