Linting#

globetrotter lint validates translation sources without writing generated files:

$ globetrotter lint
 
 
 INFO globetrotter::lint: no issues found in 0ms

The deterministic checks report:

a missing required language;
empty text or leading/trailing whitespace;
a template that does not compile with the configured engine;
a value substituted in some languages but not in others; a name that only selects a wording, such as {{#if my_condition}}, may be absent from a language without that distinction;
a condition whose branches are identical, so it has no effect;
placeholders that are undeclared, or arguments that are never used;
identical translations within one key;
duplicate text shared by different keys.

The template checks use the configured engine (or --engine) and are never run with a guessed one: without an engine, or with one that has no template analysis yet, they are skipped and a note says so. Handlebars is the only engine with template analysis today.

Findings are warnings by default and errors when strict mode is active in the config or on the command line.

Every finding carries the code shown in brackets, as in warning[duplicate]: …. Suppress one by listing it with the lint: prefix in an allow list — on the key, on an enclosing table, or on the whole config. See local lint exceptions for the scoping rules.

Find unused keys#

Pass one or more source directories to report translation keys that application code never references:

globetrotter lint --usages ./src --usages ./packages

The scanner follows ignore files and treats dynamic key prefixes conservatively. It is a cleanup tool, not a proof that every runtime-computed lookup is dead; review unused-key findings before deleting translations.

Disable duplicate detection for a run with --no-duplicates. For a deliberate exception on one key, prefer its local allow list:

allow = ["lint:duplicate"]

LLM-assisted drift review#

--llm-judge adds an experimental semantic review. It sends all languages for one key to an OpenAI-compatible endpoint and asks whether they tell the user the same thing:

# Local Ollama endpoint and the default model.
globetrotter lint --llm-judge

# A hosted or otherwise compatible endpoint.
globetrotter lint \
  --llm-judge \
  --llm-base-url https://api.example.com/v1 \
  --llm-model my-model \
  --llm-api-key-env MY_API_KEY

This review is advisory. Model findings are emitted as notes and never make lint fail by themselves; inspect the reason and the translations. The judge is deliberately tuned for recall, so false positives are expected.

Verdicts are cached by content. A rerun only judges changed keys, and changing --llm-min-confidence re-filters cached findings without sending new requests. Use --max-keys 25 to evaluate a model or prompt on a bounded slice before reviewing a large catalog.

Model choice matters. Small models can miss genuine meaning changes while inventing problems in correct translations. Evaluate a model against examples from your own catalog, keep temperature at the reproducible default, and treat reported confidence as a ranking rather than a probability.

Suppress a reviewed, intentional divergence with:

allow = ["lint:llm-drift"]

The CLI reference lists the endpoint, prompt, concurrency, effort, cache, and confidence controls.