---
name: language-detection
description: Use when identifying the source language of parsed book text and resolving the requested translation target from user instructions or configuration.
---

# Language detection for Wenyi books

Return a compact, evidence-based language decision for the parent agent. Work
from a representative sample of non-empty paragraphs, preferably from more
than one chapter. Do not require or reproduce the entire book.

## Source language

- Identify the language actually used in the sample, not the book's presumed
  country, author, title, or file name.
- Use Wenyi language codes where possible: `ja`, `zh`, `zh-Hant`, `en`, `ko`,
  `fr`, `de`, `es`, `it`, `pt`, `ru`, or `vi`.
- Distinguish Simplified Chinese (`zh`) from Traditional Chinese (`zh-Hant`).
- If the sample is mixed, quote only short evidence and state which language is
  primary. If evidence is insufficient, return `source: null` and low confidence.

## Target language

- A source-language sample alone cannot establish the intended target language.
- Resolve target from an explicit user instruction first, then a clearly
  applicable Wenyi configuration or prior translation-direction setting.
- The language used to speak to the agent is not automatically the target when
  the user asks for a different direction.
- If the target is not stated or clearly established, return `target: null` and
  ask the parent agent to clarify with the user. Do not guess.

## Response format

Return only a concise structured result:

```json
{
  "source": {
    "code": "ja",
    "label": "Japanese",
    "confidence": "high",
    "evidence": "The sample uses Japanese kana and Japanese sentence structure."
  },
  "target": {
    "code": "zh",
    "label": "Simplified Chinese",
    "confidence": "high",
    "basis": "The user explicitly requested Japanese-to-Chinese translation.",
    "digest_length": "200 characters"
  }
}
```

Use `null` for an unresolved source or target. Use `low`, `medium`, or `high`
for confidence. Never infer target language from the source text.