Localize output¶

Goal: Make a skill produce text in a chosen language without modifying the skill itself.

How language flows¶

The OS injects an output_language field into every context frame. Phase instructions that produce user-facing text honor it ("write the answer in {output_language}"). The LLM picks up the cue automatically — no per-skill localization code is needed.

Resolution order¶

--output-language CLI flag (reyn run, reyn eval, reyn chat)
reyn.local.yaml (personal override, gitignored)
reyn.yaml (project setting)
~/.reyn/config.yaml (user-global)
Built-in default (ja)

Set per project¶

# reyn.yaml
output_language: en

Every run in this project uses English unless overridden.

Override per session¶

reyn run my_skill "..." --output-language fr
reyn chat --output-language en

The override only affects that run.

Skill author guidance¶

Don't hardcode language strings in phase instructions. Reference output_language instead:

Reply in {output_language}. Use a friendly, concise tone.

The runtime substitutes the resolved value. This keeps one skill working in every language the model supports.

What this does NOT do¶

It doesn't translate input. If the user types in Japanese, the LLM sees Japanese; whether it replies in output_language or echoes the input language depends on the prompt.
It doesn't pick a model. Some languages may need a stronger model — choose with --model.
It doesn't enforce strict language output. The LLM may slip into another language under pressure (low-confidence answers, code blocks). If strict enforcement matters, add a validation step.