huangzhenpc a6e11c6d22 fix: drop --- SYSTEM PROMPT --- wrapper in Claude path to avoid Kiro injection detection ...

Kiro's upstream model is trained to identify and resist
--- SYSTEM PROMPT --- marker blocks as injection attempts, causing it
to actively reject the user's system prompt and self-correct its
identity. Switch the Claude path to the same plain-prepend approach
already used by the OpenAI path: system content is joined directly
before the user message without any marker, matching natural context.

The sanitizer (reSysPromptBlock) still strips the old marker format
from conversation history until existing contamination clears out.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-05-12 11:41:44 +08:00

36 KiB

Raw Blame History

View Raw