feat(prompt): tokenizer-style prompt stitching with thinking-prefix support

[CJackHwang]

Motivation

• Make prompt assembly follow a tokenizer/chat-template style so turn boundaries and generation-prefixes are clearer and closer to native model formats.
• Ensure assistant history and the generation prefix differ for thinking-capable models versus non-thinking models to support reasoner/chat model behavior.

Description

• Add tokenizer-style framing markers and think markers: introduce <｜begin▁of▁sentence｜>, <think>, and </think>, and apply them when assembling prompts.
• Add MessagesPrepareWithThinking(messages, thinkingEnabled) and keep MessagesPrepare as the non-thinking default wrapper for compatibility.
• When building prompts, prepend a begin-of-sentence marker, emit historical assistant content with </think>{content}<｜end▁of▁sentence｜>, and append an assistant prefix (<think> for thinking models or </think> for non-thinking models) when the last role is not assistant.
• Thread thinkingEnabled through OpenAI/Claude/Gemini prompt normalization (via deepseek) so resolved model capabilities select the appropriate suffix/prefix behavior, and update unit tests to assert the new markers and think-prefix semantics.

Testing

• Ran go test ./internal/prompt ./internal/util ./internal/adapter/openai ./internal/adapter/claude ./internal/adapter/gemini and all affected unit tests passed.
• Ran ./scripts/lint.sh, ./tests/scripts/check-refactor-line-gate.sh, ./tests/scripts/run-unit-all.sh, and npm run build --prefix webui, and these checks completed successfully.

---

Codex Task

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
ds2api	Ready	Preview, Comment	Apr 12, 2026 6:00am

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 433a3a877d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Strip think delimiters from visible model output

This change starts every generation with raw <think>/</think> delimiters, but output cleanup still only strips DeepSeek role/end markers in sanitizeLeakedOutput and does not remove think tags. When upstream echoes prompt-template tokens (the same class of leakage already handled for <｜Assistant｜>/<｜end...｜>), those delimiters will now be returned to clients as literal answer text, polluting responses and exposing internal reasoning boundaries. Please add <think>/</think> (and the new BOS marker if needed) to output sanitization or avoid emitting them into user-visible channels.

Useful? React with 👍 / 👎.