[Makar8000]

The advanced engines (GUI llama.cpp / Ollama / LM Studio advanced, seconv *-advanced) put each batch into the user message as a JSON object, {"history":[...],"lines":[{"n":1,"text":"..."}]}. That object is built with a Utf8JsonWriter on the default encoder, which writes every character above U+007F (and <, >, &, ', +) as a \uXXXX escape. The server decodes only the request's outer JSON layer, so those escapes reach the model as literal text: where the subtitle said 你昨晚在哪里？ the model read the 42 ASCII characters \u4F60\u6628\u665A\u5728\u54EA\u91CC\uFF1F.

Effect

• Every source with non-ASCII text is hit, in the GUI and in seconv alike: CJK scripts entirely, Cyrillic, Greek, Arabic, Hebrew, Thai, and Latin languages wherever they use accents or typographic punctuation. The default encoder also escapes ASCII quotes, apostrophes, angle brackets, ampersands and plus signs, so even English text with a contraction or an italics tag reached the model with escapes in it. History pairs are affected on both sides, so a Latin source with a non-Latin target loses the consistency the history is there for.
• Small and translation-tuned models cannot read the escapes. Hy-MT2 7B answered a Chinese batch with fluent English unrelated to the source ("The battle began in the dense forest." for "Where were you last night?"), and because the reply is schema-forced to the batch's line numbers, every guard in AdvancedTranslatorBase passed and the file was reported as translated. Large chat models partly decode the escapes themselves, which is why English-source tests and big models never showed it.
• The system prompt (synopsis, glossary) is written into the outer body and was decoded by the server, so it reached the model intact. That is how glossary terms turned up inside invented sentences.

Changes

• LlamaCppAdvancedProtocol gets a shared JsonWriterOptions with JavaScriptEncoder.UnsafeRelaxedJsonEscaping, used by BuildUserContent (the fix), BuildResponseFormatJson (digits only, changed for consistency) and the client's BuildRequestJson (the outer body, so captures and logs are readable and the user content is escaped once). Quotes, backslashes and control characters are still escaped; the JSON stays valid and ParseTranslations and the server-side parsers are unaffected. "Unsafe" only concerns JSON embedded in HTML, which this never is. JavaScriptEncoder.Create(UnicodeRanges.All) was not enough: it still escapes quotes, apostrophes, angle brackets, ampersands and plus signs. The relaxed encoder is what seconv's JSON output and the language files already use. Characters outside the Basic Multilingual Plane (emoji), U+2028, U+2029 and DEL stay escaped with every JavaScriptEncoder; the reply is parsed as JSON, so those come back intact in the subtitle.
• AiReviewProtocol.BuildUserContent and AiReviewClient embed the chunk JSON the same way and get the same options.
• Tests: the user content, the request body (decoded one layer, as the server does) and the AI Review chunk keep Chinese, Japanese and accented text readable with no \u escape, quotes still escaped, and the response schema still parses.

Verification

• LibUiLogicTests 1641 passed (3 new), the UI's AI Review tests 90 passed (1 new), SeConvTests 711 passed.
• seconv llamacpp-advanced, --translate-from:zh-CN --translate-to:en, llama-server with Hy-MT2 7B, an 8-line Chinese sample, requests captured through a logging proxy, same server and settings before and after:

	Before	After, 7B	After, 1.8B
\uXXXX sequences in the user message the model reads	67	0	0
Raw CJK characters the model reads	0	57	57
Requests	1	1	1
Lines translated correctly, of 8	0	8	8

Before: 你昨晚在哪里？ became "The battle began in the dense forest." and the other seven lines continued the invented battle, reported as a successful conversion. After: "Where were you last night?", "I stayed at the office until late.", "Can anyone prove it?", "The security guard saw me leave." and so on, with the 1.8B model as good as the 7B on this sample.

• Drafted with AI assistance (Claude Code); the tests and runs above were performed as listed.

🤖 Generated with Claude Code

[Copilot]

🟢 Approval recommended

No active review issues were identified in the JSON-encoding changes for preserving readable multilingual subtitle text.

No open issues 🎉

What changed in this PR

This PR keeps non-ASCII subtitle text readable in the nested JSON messages sent by advanced translation engines and AI Review, rather than exposing models to literal Unicode escape sequences.

Changes:

• Applies relaxed JSON escaping to translation batches, review chunks, and their outer request bodies.
• Reuses the translation writer options for response-schema generation.
• Adds regression tests for multilingual text, quote escaping, decoded request content, and JSON validity.

File	Description
tests/​UI/​Features/​Tools/​AiReview/​AiReviewProtocolTests.cs	Tests readable multilingual review text and actor names.
tests/​libuilogic/​Translate/​LlamaCppAdvancedProtocolEncodingTests.cs	Tests batch encoding, request content, and response-schema JSON.
src/​ui/​Features/​Tools/​AiReview/​AiReviewProtocol.cs	Defines relaxed writer options and applies them to review chunks.
src/​ui/​Features/​Tools/​AiReview/​AiReviewClient.cs	Applies review writer options to request bodies.
src/​libuilogic/​Translate/​LlamaCppAdvanced/​LlamaCppAdvancedProtocol.cs	Applies shared relaxed writer options to batches and response schemas.
src/​libuilogic/​Translate/​LlamaCppAdvanced/​LlamaCppAdvancedClient.cs	Applies translation writer options to request bodies.

_{Review effort: Balanced}