The 19-unit plan is the reference. This is the decision layer on top of it — what goes where, what runs local vs on the Paperclip team, and the five things that ship this week.
Ship and launch AINA. Every cockpit item earns its place by one test: does it help me ship faster, not lose work, or not derail? If it doesn't, it's parked. When in doubt, the answer is build Phase 1, don't plan Phase 3.
The recurring "skill vs hook vs plugin vs subagent vs team" question resolves to one rule: pick the mechanism by what fires it and where it runs.
| Mechanism | Use when | Cockpit examples |
|---|---|---|
| Hook | it MUST happen automatically, every time, regardless of the model's choice — a guarantee, not discipline | reconcile-start, closeout-stop, PreCompact handoff (U15), statusline (U1), scope-pivot (U4), retry-guard (U12) |
| Skill | a reusable way of doing X, invoked on demand or auto-matched by description — the default | prompt-scaffold (U11), landed-check (U3), which-repo, session-resume (U18), aina-build-lane, ANMS |
| Output-style | a standing persona/format for the whole session (one active) | founder-brief (U3) |
| Custom subagent | a narrow worker needs its own context + pinned tools/model — for parallelism or isolation. Never a tidiness reflex; producer ≠ verifier | verifier / fact-checker / qa-verifier (U6/U16), repo-reconciler, builder-lane, explorer |
| Agent team | many full CC instances need a shared task list + messaging — defer; Paperclip already does this | (Phase 3 comparison only) |
| Plugin | packaging skills+hooks+agents to distribute as a set (hooks.json auto-loads) | aina-agentops (adopt), ECC (evaluate) |
| Paperclip member | persistent, goal-driven, long-running, cross-lane build/launch work — the factory | Jessica → Monica/Frodo → builders → QA (Gimli/Éowyn/Calibrator) |
Decision order when unsure: no-new-machinery → skill → output-style/hook → subagent → plugin → Paperclip member. Bias down the list.
Your instinct is right — a lot of the agentops assets get passed to the Paperclip team, not run locally.
| Lane | What lives here |
|---|---|
| Local cockpit Mac Claude Code | orchestration & arbitration, planning, review councils, quick read-only exploration, all the ADHD-guardrail hooks, founder-facing surfacing (founder-brief), interactive skills (landed-check, which-repo, prompt-scaffold, session-resume), your workflow library |
| Direct Codex-over-SSH aina-build-lane | scoped, spec-frozen build work: report-file + watcher, commit-per-lane, independent verifier gate |
| Paperclip team VDS, persistent | the build/QA/release layer — ANMS spec-locks, aina-build-lane as team method, and the agentops build skills (standards-enforcer, merge-gate, dr-gate, decision-contract, verifier, codex-ops, release-core) distributed as charters to Jessica→heads→builders→QA→Frodo |
| Hermes VDS bridge | chief-of-staff bridge — being folded into native Paperclip (626-gated). Don't build new Hermes. |
codex-prompt-builder.That's it for now. U4–U19 stay parked in the plan, dependency-ordered. Building more than this in one pass is the derailment.
openclaw/skills/coding-agent (background-agent delegation → confirms the Paperclip model) and facebook/react/.claude/skills/fix (deterministic lint-before-commit → a real U3/U16 reference). I stopped at 2 pages on purpose — a 1,000-item marketplace is the exact rabbit-hole the ship guardrail refuses.
Five commits this week — a compaction backstop, an ambient statusline, ported automations, a founder-brief closeout, and a prompt scaffold. Then AINA gets your attention back.