Kimi Thinking Prefill

Client-side SillyTavern extension implementing the kimi-k3-jb patch (reasoning_content thinking prefill for Kimi/Moonshot models) without modifying any server files.

What it does

Hooks CHAT_COMPLETION_SETTINGS_READY and rewrites the outgoing request payload:

1. Injection — when the prompt does not end on an assistant message, appends:

   { "role": "assistant", "content": "", "reasoning_content": "<your prefill>", "partial": true }

   The model then continues thinking from the prefilled reasoning, per the Moonshot API's documented partial-prefill behavior.

2. Preset parity (patch transform) — if the prompt already ends on an assistant message whose content starts with <think>, the think block is moved into reasoning_content and the message is flagged partial: true — the exact transform the server patch adds to addAssistantPrefix. This means preset-style prefills like <think>I should continue the story. keep working unchanged.

3. Preserved thinking (optional) — re-attaches each prior assistant message's stored reasoning (chat[i].extra.reasoning) as reasoning_content in the outgoing payload, so Kimi's preserved-thinking behavior works in multi-turn chats. Off by default; requires the SillyTavern Show thoughts toggle to be on (reasoning is only persisted then). Note: prior reasoning is billed as input tokens.

Guards (mirroring the patch)

• Only runs when the current model id matches the configurable filter (default kimi,moonshot, covers moonshotai/kimi-k3 on OpenRouter and kimi-k3/kimi-k2-* on the direct Moonshot API).
• Skipped when JSON schema / structured output is active.
• Skipped when tools/function calling are in play.
• Injection only on normal/regenerate/swipe generations (never quiet prompts or impersonation); the <think> transform additionally applies on Continue.

Settings

Extensions menu → Kimi Thinking Prefill:

• Enable thinking prefill — toggle for the prefill features (transform + injection). Independent of the re-attach toggle below; either works without the other.
• reasoning_content prefill — the thinking text to prefill (plain text, no <think> tag needed).
• Model filter — comma-separated substrings matched against the model id.
• Force thinking on for prefilled requests (default: on) — sets include_reasoning on requests this extension modifies. Required: with thinking disabled the model continues the seeded reasoning_content with reply text and never reasons (reply shows up inside the reasoning panel).
• Send all prior assistant reasoning back to the API (default: off) — the preserved-thinking feature above. Pairs chat messages to outgoing assistant messages 1:1 (system messages excluded).
• Log decisions to browser console — debug output for each guarded decision.

Verification

Enable debug logging, generate, and check the browser console for [KimiThinkingPrefill] Injected reasoning_content prefill: .... On the server side (ST terminal with request logging), the final message should look like:

{ "role": "assistant", "content": "", "reasoning_content": "I should continue the story.", "partial": true }

Notes

• Works with the direct Moonshot source, OpenRouter (Moonshot provider), and Custom endpoints — anywhere the model id matches the filter and the API honors partial/reasoning_content.
• The server patch is not required; do not run both (double transforms are harmless but pointless).