Repository navigation
Releases
Releases · Trapa-Eureka/live-skill
v0.1.2
Documentation-only release, published to npm as live-skills@0.1.2, so the npm page matches the repository README. No code changes since 0.1.1.
Changed
- README restructured: a Features section, a "Where live-skills can be used" section organized by field (collapsible groups), the 60-second demo section removed, and the Status section reduced to the essentials. The Korean README mirrors the same structure.
Full notes: CHANGELOG.md
v0.1.1
Patch release, published to npm as live-skills@0.1.1. The first real-LLM runs after 0.1.0 exposed three defects that made compile fail on real input (0.1.0 fails at the outline step; upgrade).
Fixed
- Outline and qaGen responses wrapped in a ```json code fence or surrounded by a sentence are now parsed (envelope stripped; the strict schema check is unchanged).
- Section ids copied with the § marker from the prompt header (
§overview,[§overview]) are normalized to the bare id before the coverage check, and the outline prompt asks for bare ids. - Golden-question anchors are the shortest fact-bearing span instead of whole sentences, and the distillation prompt keeps figures, units, addresses, identifiers, commands, and component names verbatim, so distilled chapters no longer fail the verbatim anchor check merely by paraphrasing.
outline_invalidfailures name the reason (not valid JSON (…)orschema: <path>: <rule>).
Full notes: CHANGELOG.md
v0.1.0
First public release, published to npm as live-skills.
Added
compile: text PDF, DOCX, Markdown/TXT, and HTML sources (files, folders, globs) into an Agent Skills-standard skill:SKILL.mdindex, chapters with source anchors,glossary.md,patterns.md,cheatsheet.md, and amanifest.jsonwith source and output hashes.- Quality gate: golden Q&A generated from the source, answered using only the compiled skill files, graded conservatively; deployed only when the pass rate meets the threshold (0.9 by default, floor 0.5), otherwise kept in a temp directory with a report naming the weak chapters.
validate(structure check, no LLM calls),report(last gate report),eval(re-grade an existing skill, optionally against changed sources).- Cost and safety limits:
MAX_LLM_CALLSenforced during the run, input size limits, parser timeouts, prompt data/instruction boundaries, stale/tampered detection for compiled output. - Claude as the default provider via
ANTHROPIC_API_KEY; the test suite runs against a scripted mock LLM with no network access.
Full notes: CHANGELOG.md