ohyeh/agent-scripts

shared-memory-intake

Curate shared Codex memory when Claude submits findings or Codex reviews, promotes, or maintains task memory.

소스 보기

사용 방법

  1. 설치 명령을 복사하고 자신의 에이전트 환경에 설치하기 전에 소스 패키지를 확인하세요.
  2. 필요한 도구, 인증 정보, 지원 환경은 원본 문서에서 확인하세요.
  3. 입력과 원하는 산출물을 설명하고 출처와 결과를 대조하세요.
npx skills add https://github.com/ohyeh/agent-scripts --skill shared-memory-intake

시작 전 확인

  • 설치에는 모델 구독, API 한도, 파일 접근 권한이 포함되지 않습니다.
  • 별점 수와 설치 수는 활동 지표이며 독립적인 품질·보안 검증이 아닙니다.
원본 Skill 문서

원본 저장소의 제목, 예시, 코드, 표, 링크, 이미지를 유지해 표시합니다.

Shared Memory Intake

Use ~/.codex/memories/ as the canonical shared knowledge store and ~/.agents/shared-memory-inbox/ as the external submission store. Read MEMORY.md before opening matching rollout_summaries/ files. Treat dynamic claims as UNCONFIRMED until verified in the active session.

Actor contract

Require an explicit actor: --as claude or --as codex. Never infer it from the runtime.

ActorAllowed operationsForbidden operations
clauderead official memory; create one submission in inbox/pending/edit or delete official memory, or another submission
codexread, review, promote, supersede, maintaindelete historical summaries or silently overwrite a current claim

Claude: submit

  1. Read MEMORY.md; do not restate an existing current claim as a new submission.
  2. Create one uniquely named Markdown file under ~/.agents/shared-memory-inbox/pending/ with YAML frontmatter:
yaml
   schema_version: 1
   source_runtime: claude
   submitted_at: 2026-07-26T00:00:00Z
   cwd: /absolute/path
   source_session: opaque-session-id
   claim_status: unverified
   evidence_host: hostname-or-fleet-id

evidence_host is optional for same-host evidence and required when the cited source exists only on another host.

  1. Keep only reusable conclusions, source references, evidence, and unknowns. Never

include credentials, tokens, private keys, cookies, or raw transcripts.

  1. Validate before handoff:
sh
   python3 ~/.agents/skills/shared-memory-intake/scripts/validate_submission.py /absolute/path/to/submission.md
  1. A correction is a new submission with supersedes_submission; never overwrite the

earlier proposal.

Codex: review and promote

  1. Acquire the single-curator lease before promotion: mkdir ~/.agents/shared-memory-inbox/.curator.lock.

If it already exists, stop and report the active-curator conflict. Release it with rmdir after the outcome is recorded.

  1. Run the validator on the candidate. Reject on schema or secret-pattern failure.
  2. Inspect the cited source and live-verify every dynamic claim that would enter the

current index. For cross-host evidence, require evidence_host, verify the live source on that host, and record the curator host plus verification time in the promoted summary. Missing host access remains UNCONFIRMED.

  1. Re-read MEMORY.md immediately before its update. If its relevant topic changed

during review, restart the review instead of merging assumptions.

  1. Promote accepted knowledge to a new immutable rollout_summaries/*.md record and

update MEMORY.md as the only current index. Give every changed reusable claim a lifecycle state: active, superseded, or UNCONFIRMED.

  1. Move the candidate to processed/ or rejected/ without deleting its audit trail.

Codex: maintain

  1. Acquire the same single-curator lease and re-read MEMORY.md immediately before

every official-memory write; stop on an active-curator conflict.

  1. Scan current index entries for changed paths, versions, models, deployments, or

stale references.

  1. Demote unverified dynamic claims to UNCONFIRMED; mark replaced claims

superseded with their replacement reference.

  1. Consolidate duplicate current entries. Do not rewrite historical rollout summaries.
  2. Re-run the validator on any newly promoted submission and report the paths changed.

Validator

Use scripts/validate_submission.py. It is a narrow preflight gate, not proof that a claim is true and not a complete secret scanner. A pass authorizes review only; Codex still owns evidence verification and promotion.

관련 작업

같은 저장소의 Skills

더 많은 Skills

모든 Skills
ohyeh
커뮤니티

using-design-skills

Which design skill, in what order. Invoke BEFORE any visual output: web page, landing page, dashboard, product UI, mobile screen, HTML report, diagram, chart, plan page, artifact, motion, redesign, 「做個頁面」, 「畫圖」, 「報告做成網頁」, plus module or API interface design. Picks one direction authority and an executor per stage, closes with an evidence check. Even a 1% chance means invoke it.

1시간 설치 수
1
GitHub Stars
0
업데이트
10월 5일
ohyeh
커뮤니티

ask-nova

Interview-driven flow picker. Use when the user is unsure which skill/flow/pipeline fits — new project kickoff, cross-domain work, or a fuzzy situation. Reads the three routers live and recommends ONE flow. Not for tasks whose owner is already obvious.

1시간 설치 수
3
GitHub Stars
0
업데이트
9월 2일
ohyeh
커뮤니티

defect-first-review

Perform a read-only, defect-first review of a specified code change and return every actionable finding, reading the whole diff before running anything. Use when another agent delegates review of uncommitted changes, a base-branch diff, a commit, or custom review instructions. Prefer this over the Codex-bundled review-agent, which carries no read-before-you-run gate.

1시간 설치 수
3
GitHub Stars
0
업데이트
9월 2일
ohyeh
커뮤니티

delegation-templates

Fill-in-the-blank prompt templates for delegating work to subagents or tmux workers - SEARCH/LOCATE, IMPLEMENT, REFACTOR, RESEARCH, REVIEW/VERIFY. Invoke BEFORE writing any delegation prompt (Agent tool, agent-tmux worker, or fanout task) so the prompt ships with explicit GOAL, ACCEPTANCE, and REPORT sections instead of a vague ask. Not for deciding WHETHER to delegate or for driving workers after launch (see the tmux-delegate agent and the tmux-agent-tools skill for those).

1시간 설치 수
3
GitHub Stars
0
업데이트
9월 2일