Development Log: Project Claim Output Brake

Purpose

Stop unsupported project technical claims in the final answer body, not only in Trace diagnostics.

Problem

Even when Trace reported projectClaimPolicy=general-only, the model could still write phrases such as:

"기술적 기반 면에서는 안정적"
"아키텍처는 유연하나"
"모듈화된 구조는 향후 기능 추가 시 유연성을 제공합니다"

Prompt guidance alone was not enough.

Implementation Summary

Added enforceProjectClaimPolicyInAnswer.
When projectClaimPolicy=general-only, final answer text is scanned before output.
Unsupported technical structure claim sentences are removed.
A safe notice is inserted:
- "현재 정보만으로는 기술 구조를 판단할 수 없습니다."
- "기술적 안정성, 아키텍처 유연성, 모듈화 여부는 소스 코드나 설계 문서 확인이 필요합니다."
Connected the enforcement function in AgentExecutor before final response streaming and history persistence.
Added regression tests for the exact problematic phrases.

Changed Files

src/features/secondBrainTrace.ts
src/agent.ts
tests/secondBrainTrace.test.ts

Verification

./node_modules/.bin/tsc --noEmit
npm run compile
./node_modules/.bin/jest --runInBand

Result

The system now has a deterministic output brake for unsupported project implementation claims when Trace says only general/reference evidence is available.

1.5 KiB Raw Blame History