Skip to main content
Version: Next (unreleased)

cl10n — continuous localization for Markdown

Why it exists​

The same problem as cedit, one step over: translations. A skill is a Markdown document, and once it is translated, every upstream change puts the mirror out of date. Running an LLM over the whole file again is the obvious move and the wrong one — it costs the full document every time, and it re-words paragraphs nobody touched, so the diff of the translation is unreadable.

cl10n retranslates only what actually changed. Everything else is served from a committed translation memory.

How it works​

Each document is parsed into an AST and Merkle-hashed, and two revisions are diffed structurally. The result is a per-paragraph verdict — reuse, recheck, revise, translate — and only the last two become LLM jobs. Rendering then splices the translations back into the source tree, so headings, list nesting, table shape and code fences come from the original and a translation cannot corrupt them.

pip install markdown-localization[groq]   # also [nvidia], [mistral], [all-providers]
export GROQ_API_KEY=...

cl10n plan   --langs he,ru                # diff the corpus into a queue of jobs
cl10n run    l10n/queue/queue.json -c 8   # execute the queue
cl10n render --langs he,ru                # memory → locales/he/**, locales/ru/**
cl10n status --langs he,ru                # coverage per language

The distribution is markdown-localization; the command and the importable package are both cl10n. Python 3.11+, and providers are pluggable.

The same four commands serve a first translation and a daily update — a first translation is just an incremental update whose previous revision was empty. Kill a run and re-run it: the resume state is the memory rather than the queue, so finished work is never re-billed and interrupted work is never lost.

What the pipeline guarantees, and which the plugin's own docs would need: a paragraph repeated in two files costs one translation; inline code, link targets and image sources are checked against every response, at translation time and again at render time; a render refuses to write a file whose block structure moved; and a unit with no usable translation falls back to English and is counted, never shipped silently.