Ardig24/dsh-trajectory-ablation0

dsh-trajectory-ablation

Finds the actual cause of an agent failure by reconstructing, diffing, and ablating the context blob the model saw at each step. Plugin for DeepSeek Harness.

AI 분석

核心用于对智能体(Agent)的决策失败进行深度 Debug。适合智能体开发者,用于分析上下文冗余或提示词干扰问题。

패키지
dsh-trajectory-ablation
버전
0.1.0
라이선스
MIT
최근 업데이트
2026. 8. 25.

설치

검증된 bundle이 없거나 호환성 검사에 실패했습니다. 먼저 저장소 설명을 읽어 주세요. 전체 README 읽기 ↗

Usage

Once mounted, three commands are available in any session:

/context-reconstruct              — what did the model see before this step?
/context-diff     — what changed between two steps?
/ablate   [k]                      — which block actually caused this decision?

Or use it programmatically — reconstructContext and diffContext need only a session's event log (no live model connection required); ablateStep needs a live ctx.llm:

import { reconstructContext, diffContext, ablateStep } from 'dsh-trajectory-ablation'