Lithuanian Text Normalizer

skills.sh

An experimental, vendor-neutral Agent Skill that defines how an agent may convert written Lithuanian into context-appropriate spoken-form Lithuanian. New expansions remain unaccented, while stress already present in the input is preserved exactly for pipelines that perform stress restoration separately.

Important

This repository currently contains agent instructions and a validation protocol, not a deterministic Lithuanian normalization engine. Unsupported or ambiguous spans must be preserved and reported. Do not use agent output as corpus ground truth.

Repository layout

plugin.json                              portable Agent Plugin manifest
skills/lt-text-normalizer/               canonical Agent Skill
  references/examples.md                 preserve/expand/abstain examples
  references/publication-examples.md     attributed literature examples
  references/published-rule-catalog.md   all 60 published rule blocks
  references/published-rule-families/    class router and per-rule examples
  references/implementation-landscape.md public-code provenance and limits
.codex-plugin/plugin.json                Codex compatibility manifest
.claude-plugin/plugin.json               Claude Code plugin manifest
.claude-plugin/marketplace.json          Claude Code marketplace entry
docs/EVIDENCE.md                         publication alignment and limitations
docs/VALIDATION.md                       behavioral evaluation protocol
scripts/validate_package.py              dependency-free package checks

The canonical behavior lives only in skills/lt-text-normalizer/SKILL.md. Vendor manifests package that same skill; they must not fork its linguistic rules.

Use

Give the installed skill Lithuanian text directly:

Normalize this Lithuanian text: Turiu 2 knygas.

For a fully resolved input, the skill returns only:

Turiu dvi knygas.

No model identifier or normalization declaration is required. The skill accepts uppercase, lowercase, mixed-case, stressed, partially stressed, and unstressed input. Ambiguous spans remain unchanged and are reported separately.

Installation

skills.sh

Install the skill for a supported agent directly from the public repository:

npx skills add MELDYNAMICS/lt-text-normalizer

Agent Skills-compatible hosts

Clone this private repository, then install or copy skills/lt-text-normalizer/ into the host's skills directory. The folder follows the Agent Skills specification.

Codex and ChatGPT

The repository root is a portable Agent Plugin and also contains a Codex compatibility manifest. Add the repository through the host's repository or local plugin flow. The package layout follows OpenAI's plugin packaging documentation.

Claude Code

Add the public GitHub repository as a Claude Code marketplace, then install the plugin:

/plugin marketplace add MELDYNAMICS/lt-text-normalizer
/plugin install lt-text-normalizer@meldynamics-lt-language-tools

Claude can select the skill automatically from a normalization request. To invoke it explicitly, use /lt-text-normalizer:lt-text-normalizer followed by the Lithuanian text. The same skill may also be copied directly into a Claude skills directory.

Hermes and other agents

Clone the repository and copy skills/lt-text-normalizer/ to the agent's Agent Skills directory. For Hermes, that is normally ~/.hermes/skills/lt-text-normalizer/.

Validate

python3 scripts/validate_package.py

Before any production claim, run the behavioral protocol in docs/VALIDATION.md against reviewed Lithuanian gold fixtures and publish immutable agent, skill, and fixture revisions. Source-linked, non-gold review candidates are available in tests/literature-candidates.jsonl.

See the skill's behavioral examples for direct normalization, casing, ambiguity, Unicode stress, and output patterns. The separate publication examples retain publication provenance and limitations. The published-rule catalog maps every numbered block (1)–(60) in the paper and explicitly records the unavailable remainder of its reported implementation. The rule router splits those rules into four class branches and gives every rule a paper or clearly marked schematic example plus a usage guard.

For implementation work, the implementation landscape records pinned public LIEPA-derived code, rule-asset hashes, license signals, and why those projects must not be represented as the paper's exact system.

Research and corpus boundary

docs/EVIDENCE.md records what the current specification does and does not implement from the referenced literature. Corpus-derived and synthetic training artifacts remain governed by the separate lt-stress-and-normalization-corpus project. Runtime acceptance does not authorize reusing an agent's output as a training label.

License

This work is licensed under the Creative Commons Attribution 4.0 International License. When sharing or adapting it, credit “MELDYNAMICS — Lithuanian Text Normalizer,” link to this repository and the license, and indicate material changes.