[scarletkc]

Updates .agents/skills/typesafe-ai/SKILL.md from upstream v0.5.7 to v0.6.2 (typesafe-ai/skills@302eea6, released 2026-10-23). The file matches upstream byte for byte. LICENSE is unchanged.

What's new in the skill

• Box question type: locate something in an image, alongside Choice, Noul and Score.
• Judgments from Jev, decisions in code: treat each decision as a truth table. Jev answers the judgments it depends on, and code maps them to an action. Asking Jev "should this be escalated?" is flagged as the wrong layer.
• New guidance sections:
  • designing state, using the "human test": could a person answer from this state at a glance?
  • writing questions and criteria, including a companion Noul for the null case
  • reading probabilities, with 0.7 / 0.3 as starting Noul thresholds and confidence rarely used as the gate
  • testing against the live API, which requires asking the user first
  • an ordered checklist for wrong answers
  • "yellow flags" for over-fitted wording
• Doc links move to the docs.typesafe.ai/go/... paths, and links are added for the OpenAPI spec and the jaggedness page.

Lock file

skills-lock.json gets the new computedHash. I computed it with the skills CLI's computeSkillFolderHash: sha256 over each file's relative path followed by its contents, sorted by path. The previous hash was produced from a CRLF checkout (Git for Windows, core.autocrlf=true). Recomputing the old version the same way reproduces it exactly, so the new hash uses the same CRLF basis. On a Linux machine npx skills update would compute an LF-based hash, see a mismatch, and reinstall the skill. The same was already true before this change.

Testing

Documentation and lock-file change only; no code paths touched.