v2.3.0 — distribution, CI, deeper OpenAI strategy

Distribution, automation, and a deeper OpenAI strategy. Routing and runtime behavior are unchanged.

Added

• Published to npm as prompt-refine-skill (files whitelist + publishConfig pinned to the official registry).
• CI: dependency-free strategy validator, plus tag-triggered npm publish with provenance.
• ROADMAP.md, a Limitations section in the READMEs, npm version/downloads badges, and a star CTA.

Changed

• OpenAI strategy deepened for the GPT-5 family (5.1 / 5.2 / 5.5) — outcome-first prompting, reasoning effort, agentic stopping/persistence/verification, and a Codex section (verify-your-work, decompose, durable rules in AGENTS.md).
• New Codex before/after example; example #2 modernized to GPT-5 (Thinking).
• Routing labels in SKILL.md made consistent with the READMEs (OpenAI → GPT-5; DeepSeek → V4).

First release published automatically via the new GitHub Actions pipeline.

See CHANGELOG.md.

v2.2.0 — accuracy & repository-health

Accuracy and repository-health release. Strategy design and routing are unchanged.

Changed

• Vendor strategies aligned with each vendor's current official prompt-engineering docs (corrected source links and claims).
• Refreshed for current model lines: OpenAI GPT-5 family; DeepSeek V4 main line (V4-Pro / V4-Flash, dual Thinking/Non-Thinking modes, 1M context) with R1 kept as a legacy, R1-scoped case; Cohere Command A family.

Added

• Multimodal (vision) framing rule across the vision-capable strategies plus the universal fallback.
• Repository governance: SECURITY.md, CODE_OF_CONDUCT.md, issue/PR templates; Private Vulnerability Reporting enabled.

Note: the README Evaluation numbers describe 2.1.0 behavior and were not re-run for this release. The changes here are accuracy/governance, not a behavioral redesign.

See CHANGELOG.md for details.