Personal Project

oh-my-field: evidence-backed agent capability platform

oh-my-field: evidence-backed agent capability platform

외부 에이전트가 한 번 잘 수행한 작업을 evidence, harness, portability 상태를 가진 재사용 가능한 capability package로 바꾸는 개인 프로젝트입니다.

A personal project that turns useful external-agent runs into reusable capability packages with evidence, harnesses, and portability state.

CLI + MCPEvidenceHarnessPortability
oh-my-field product flow from agent runtimes to portable capability packages

문제

Problem

강한 agent run에는 목표, 판단 기준, 안전한 명령, 테스트 결과, 실패 이력, 사용자 피드백이 함께 들어 있습니다. 하지만 대부분의 지식은 채팅 기록이나 터미널 로그에 남고, 다음 프로젝트나 다른 런타임에서 재사용하기 어렵습니다.

Strong agent runs contain goals, judgment criteria, safe commands, tests, failures, and human feedback. Too much of that knowledge stays trapped in chat history or terminal logs, making it hard to reuse in another project or runtime.

설계

Design

  • Codex, Claude Code, Hermes, generic/local agent가 실제 작업을 수행하고, OMF는 그 결과를 evidence record로 구조화합니다.
  • Codex, Claude Code, Hermes, and generic/local agents do the work; OMF structures their outputs into evidence records.
  • import-run과 session tracking으로 logs, patches, test results, command output, human review를 수집합니다.
  • import-run and session tracking collect logs, patches, test results, command output, and human review.
  • promote는 evidence를 runtime-neutral capability.yaml, instructions.md, harness.yaml, capability card로 승격합니다.
  • promote turns evidence into runtime-neutral capability.yaml, instructions.md, harness.yaml, and a capability card.
  • exported, imported, validated, portable 상태를 분리해 단순 변환과 실제 target-side 검증을 구분합니다.
  • Exported, imported, validated, and portable states stay separate so format conversion is not confused with proven target-side reuse.

기여

Contribution

제품 개념, CLI 명령 표면, MCP tool surface, artifact schema, runtime adapter, 안전한 command approval 정책, capability health/verify/eval/review/learn 흐름, 문서와 공개 사이트까지 직접 설계하고 구현했습니다.

Designed and implemented the product concept, CLI surface, MCP tool surface, artifact schemas, runtime adapters, command approval policy, capability health/verify/eval/review/learn flows, documentation, and public project site.

Implementation

핵심 구현 범위

Implementation Scope

Artifact Pipeline

agent run을 evidence로 import하고, capability package로 승격한 뒤 health, replay, eval, review로 검증합니다.

Imports agent runs as evidence, promotes them into capability packages, then verifies them with health, replay, eval, and review flows.

Runtime Portability

Codex, Claude Code, Hermes, generic runtime adapter를 두고 export/import/validate lifecycle을 분리했습니다.

Uses Codex, Claude Code, Hermes, and generic runtime adapters with separated export/import/validate lifecycle state.

Safety & Learning

위험 명령 승인, secret redaction, learning patch, dataset export를 통해 자동화와 검토 가능성을 함께 유지합니다.

Keeps automation reviewable through risky-command approval, secret redaction, learning patches, and dataset export.