neuron-explorer
Qwen3.5‑4B 전용 조향형 뉴런 탐색기, 모든 MLP 뉴런을 밀어 결과를 보관·탐색
AI 에이전트 개발을 위한 저장소 표준으로, 독립 리뷰 아이덴티티·실패 차단 게이트·레포 간 전파를 제공합니다
A repository standard for AI-agent development: independent review identities, fail-closed gates, and cross-repo propagation—with the failures on the record.
개발 도구Shell · PythonGitHub @nathanjohnpayne온라인운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
개인 계정의 공개 저장소 소유자는 GitHub로 로그인해 운영자를 확인하고 소개를 직접 관리할 수 있습니다. 관리자 승인은 필요하지 않습니다.
로그인한 뒤 확인 파일을 만들어 이 저장소의 기본 브랜치에 30분 안에 커밋해 주세요. 다른 브랜치나 포크의 파일은 인정하지 않습니다. 파일을 올린 뒤 아래에서 확인하면 관리자 승인 없이 관리할 수 있습니다. 자격을 유지하려면 파일을 남겨 두세요.
로그인이 필요하면 로그인 화면으로 이동합니다. 로그인 후 이곳으로 돌아와 원하는 확인 버튼을 다시 눌러 주세요.
프로젝트 폴더에서 /nomorevibe verify를 실행하면 소유권을 확인하고 소개를 직접 관리할 수 있습니다.
1분 만에 확인하기A repository standard that makes AI coding agents reviewable: canonical docs every agent must read, 70+ fail-closed CI checks, multi-identity review so no agent approves the PR it wrote, automated external review, and one-command propagation to every downstream repo.
저장소 README에서 가져옴
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
Reference implementation of the AI Agent Tooling Standard.
The goal is to allow multiple AI coding agents and development tools to operate consistently without configuration drift. See BRAND.md for the umbrella vocabulary (Playground, Cockpit, Tiebreaker, Checks) and naming history.
Read these files in order before taking any action:
1. AGENTS.md — behavioral instructions and operating rules
2. rules/repo_rules.md — binding structural constraints
3. Relevant specs/ files — intended behavior
4. .ai_context.md — supplemental system contextREADME 전문 보기 ↗Reference implementation of the AI Agent Tooling Standard.
The goal is to allow multiple AI coding agents and development tools to operate consistently without configuration drift. See BRAND.md for the umbrella vocabulary (Playground, Cockpit, Tiebreaker, Checks) and naming history.
Read these files in order before taking any action:
AGENTS.md — behavioral instructions and operating rulesrules/repo_rules.md — binding structural constraintsspecs/ files — intended behavior.ai_context.md — supplemental system contextEvery change in this repository goes through the policy in REVIEW_POLICY.md, including a self-peer review by the authoring agent's reviewer identity and, for changes that cross the threshold or touch protected paths, automated external review via the OpenAI Codex GitHub app (Phase 4a) or a manual CLI fallback (Phase 4b).
| File | Purpose |
|---|---|
AGENTS.md | Instructions for AI agents |
DEPLOYMENT.md | Build and deployment |
CONTRIBUTING.md | Development workflow |
.ai_context.md | High-level system context |
BRAND.md | Mergepath umbrella vocabulary (surfaces, reserved names, naming history) |
mergepath/playground/index.html | Mergepath Playground — tune the review policy and replay recent PRs against the draft |
scripts/policy-sim.sh | Bakes real gh PR data into a temp copy of the Mergepath Playground for local replay |
scripts/cockpit.sh | Launches the Mergepath Cockpit, a local browser dashboard over the fleet's PRs, CI and agent runs |
scripts/worktree-cleanup.sh | Audits stale git worktrees and, with --apply, removes the safe ones |
scripts/install-worktree-cleanup-agent.sh | Installs the daily launchd job that runs the worktree cleanup unattended |
.mergepath-sync.yml | Propagation manifest for synced canonical, kit, and templated surfaces |
.github/ISSUE_TEMPLATE/ | Synced issue-template kit; consumers may add product-specific templates |
.github/pull_request_template.md | Synced canonical pull-request template |
ai_agent_tooling_standard.md | Full repository standard (reference) |
These run on your own machine and never in CI.
Mergepath Cockpit is a local browser dashboard. Warm the credential cache, then launch it:
eval "$(scripts/op-preflight.sh --agent cursor --mode review)"
scripts/cockpit.sh
Agent history reads at most 64 checkouts: the hub and any configured roots first, then the newest worktrees. When there are more, Cockpit says so on startup (Agent history reads 64 of N checkouts…) and the history panel shows coverage as incomplete. If launch fails, the message names the step that failed, for example Cockpit cannot start: failed while discovering checkouts and worktrees. Behavior is specified in specs/cockpit_foundation.md and specs/cockpit_agents.md.
Worktree cleanup removes worktrees left behind by agent sessions. A dry run lists what it would do; --apply removes only safe candidates, meaning the upstream branch is gone or the PR is closed or merged, and nothing in the working tree would be lost (uncommitted, untracked or gitignored files such as .env):
scripts/worktree-cleanup.sh # dry run: report only
scripts/worktree-cleanup.sh --apply --min-age 24h # remove safe candidates untouched for a day
Locked worktrees and orphan directories are left alone unless you pass --force-locked or --orphan-clean. A worktree git refuses to remove is reported under failed: and the run carries on. Committed work survives: removing a worktree keeps its branch, and a detached worktree is removed only when a local branch, tag or remote contains its commit or GitHub still holds it as the PR's head. A detached worktree holding a commit nothing else has is kept, with the command to save it.
Daily cleanup. To run that cleanup automatically, install the per-user launchd job from a clone whose main checkout is on an up-to-date main:
scripts/install-worktree-cleanup-agent.sh # install; runs daily at 04:17
scripts/install-worktree-cleanup-agent.sh --check # installed, loaded and current?
scripts/install-worktree-cleanup-agent.sh --run-now
scripts/install-worktree-cleanup-agent.sh --uninstall
The job runs the main checkout's scripts/worktree-cleanup.sh --apply --min-age 24h and appends to ~/.cache/mergepath/worktree-cleanup.log. It never passes --force-locked or --orphan-clean. Details are in docs/agents/operating-rules.md and specs/worktree_cleanup_audit.md.
The template retains the approved 1Password-backed Firebase credential-custody model: each project's deployer service-account key belongs in the Firebase vault, with shared 1Password ADC retained as a legacy resolver fallback. Do not revert to deploy keys on disk or routine firebase login guidance without explicit human approval. Existing op-firebase-setup and op-firebase-deploy helpers describe legacy account setup and credential resolution; do not run them to enable production deployment while the production block remains.
Production deployment is blocked until separately reviewed tooling ensures repository build code and dependencies have neither deploy credentials nor author/reviewer PATs and cannot reach credential-manager files or preflight caches; publishes an immutable artifact with a digest and attestation bound to the intended repository and full source revision; and has trusted post-build tooling verify that binding and artifact integrity before acquiring deploy credentials, then deploy the verified artifact without rebuilding or executing the candidate checkout and inject each credential only into its consuming subprocess. The current helper scripts do not provide these guarantees. Do not use deploy/all preflight or legacy helper commands to bypass the block; see DEPLOYMENT.md and Build and deploy isolation.
The scripts/gcloud/gcloud wrapper remains available for ordinary non-deployment gcloud commands. Routine replacement of a Firebase deploy key remains incomplete until separately reviewed, non-deploying validation proves the replacement usable; suspected-compromise containment is a separate owner-led incident decision. See key rotation guidance.
| Directory | Purpose |
|---|---|
rules/ | Binding repository constraints |
specs/ | Intended system behavior |
plans/ | Execution and migration plans |
mergepath/ | Mergepath Playground and reserved slots for future Mergepath surfaces (see BRAND.md) |
tests/ | Automated validation |
src/ | Application code |
functions/ | Backend handlers |
scripts/ | Build, CI, and automation tooling |
docs/ | Architecture and design documentation |
dist/ | Generated build artifacts (do not edit manually) |