Ceremony

agent-tooling

Shared ways-of-working for the WEDDEsign intelligence stack — Mentra, VectraIQ, intelligence-core, and NordScope.

This repo is the single source of truth for the ceremony every repo runs: the Codex review autopilot, PR/issue-link discipline, the task-routing labels, and the security-surface guard. Write a rule once here; every repo inherits it.

Why this exists

The four repos share an identical workflow — GitHub PRs, Codex review rounds, Closes #N issue-link discipline, task-routing labels — but that knowledge lived as passive prose duplicated across four CLAUDE.md files (NordScope 640 lines / 61 codex mentions → VectraIQ 174 / 0). It was re-interpreted every session, drifted between repos, and was not executable: nothing enforced it.

This repo makes the ceremony executable and shared. The guiding rule (from the toolbox-thinking model that kicked this off): rank tooling by frequency × friction removed, and prefer boring infrastructure over flashy demos. Every asset here is high-frequency ceremony, so it is shared-by-default.

The two planes

Ceremony runs on two planes. You need both — neither covers the other's cases.

Plane	Vehicle	Fires when	Covers
Client-side (sync)	The ceremony plugin — skills, commands, hooks	A human is in an interactive Claude Code session	Guidance the agent follows + guards it cannot skip locally
Server-side (async)	Reusable GitHub workflows in .github/workflows/	No session is open — Codex finishes a review, CI turns green/red	Advancing the PR loop while nobody is watching

The plugin cannot advance a PR when no session is open; the workflows cannot scaffold a page or run an eval before push. Ship both.

Layout

agent-tooling/
├── .claude-plugin/marketplace.json     # catalog (one marketplace, one+ plugins)
├── plugins/
│   └── ceremony/                        # the shared client-side plugin
│       ├── .claude-plugin/plugin.json
│       ├── skills/
│       │   ├── codex-review-triage/     # the Codex autopilot loop (extracted from NordScope)
│       │   └── open-pr/                 # Closes #N + test-plan discipline
│       ├── commands/                    # /delegate-to-codex, /fanout
│       └── hooks/                       # security-surface-guard (PreToolUse)
├── .github/workflows/                   # reusable (workflow_call) ceremony — the server-side plane
│   ├── wake-on-codex-review.yml
│   ├── wake-on-ci-green.yml
│   ├── wake-on-ci-red.yml
│   ├── auto-resolve-review-threads.yml
│   └── cleanup-stale-codex-labels.yml
├── templates/caller-workflows/          # thin per-repo callers that invoke the reusable workflows
├── labels/                              # canonical label taxonomy + gh sync script
└── docs/per-repo-wiring.md              # how each repo opts in

Install (per consuming repo)

1. Labels (one-time): labels/sync-labels.sh WEDDEsign/<repo>
2. Plugin: commit .claude/settings.json with the marketplace + enabledPlugins (see docs/per-repo-wiring.md).
3. Workflows: copy the matching templates/caller-workflows/* into the repo's .github/workflows/, set the repo-specific check names, add the CEREMONY_PAT secret.

Full per-repo steps and the VectraIQ / intelligence-core / NordScope deltas are in docs/per-repo-wiring.md.

Governance

• Promotion rule: domain tooling stays repo-local; ceremony is shared-by-default; a local skill graduates here on its 2nd consumer.
• Source of truth is NordScope's proven loop — the autopilot here was extracted from it, not invented. Improve it here, not in a fork.
• Definition of done (a shared asset): used in real work ≥1×; no hardcoded repo specifics (everything repo-specific is an input/config); namespaced.
• Precedence hygiene: shared skills are namespaced ceremony:<skill>; never create a same-named skill in a repo's local .claude/.
• Ownership: one owner for versioning/releases of this repo.

Dogfood before publish

Per plan: prove this locally before pushing to GitHub. Add it as a local-path marketplace — /plugin marketplace add ./agent-tooling — wire the Mentra pilot against it, run a real PR through the loop, then publish to WEDDEsign/agent-tooling and flip repos to the GitHub source.

Code & architecture doctrine

doctrine/CODE_DOCTRINE.md is the canonical quality/maintainability/scalability standard Codex and Claude hold every change to. Enforced three ways: the doctrine/file-size-ratchet.py ("don't get worse" gate, run in preflight + CI), Codex review (the doctrine is synced into each repo's AGENTS.md via doctrine/sync-doctrine.py), and Claude while writing. Universal rules live in the doctrine; repo-specific patterns stay in each repo's AGENTS.md.