Ideation: Move Brainforge agent assets into the team plugins marketplace

Date: 2026-04-22
Focus: /ce-ideate — how to relocate (or publish) Brainforge-specific agent assets—Cursor skills, rules, command patterns, MCP templates, and related guidance—so they travel with engineers across repos and workspaces, not only when brainforge-platform is open.
Mode: Repo-grounded — assets today live primarily in brainforge-platform (.cursor/, AGENTS.md, .agents/ for Codex, opencode.jsonc, sync scripts); distribution already started in brainforge-ai/brainforge-plugin-marketplace with per-vendor plugin folders.
Related plans: docs/plans/2026-04-21-002-feat-cursor-team-marketplace-repo-plan.md (dedicated marketplace repo; monorepo documents, does not replace). This ideation extends that with what to ship inside marketplace plugins vs what stays in the monorepo.

Session note: Full ce:ideate multi-agent dispatch was not run; this artifact follows the skill’s structure (grounding → many candidates → explicit rejection → ranked survivors) in one pass over AGENTS.md, .cursor/ layout, and marketplace state.


1. Grounding summary

What “agent assets” means here

BucketWhere it lives todayPortable without monorepo clone?
Cursor skills.cursor/skills/ (many; some git submodules via scripts/ensure-git-submodules.mjs)No — repo-scoped unless mirrored to user or marketplace
Cursor rules.cursor/rules/*.mdcNo — project rules follow the repo
Compound Engineering.cursor/settings.json + optional bunx @every-env/compound-plugin for OpenCode/CodexPartially — different install paths per surface
Codex skills.agents/skills/ + user ~/.agents/skills/User path is portable; repo path is not
“Cursor mirror” to Codexsync-cursor-skills-to-codex.py (user-level mirror)Yes for individuals who run the script
MCP.cursor/mcp.json, docs, per-server OAuthTemplates portable; secrets and OAuth stay per machine
OpenCodeopencode.jsonc, user config mergeMostly user + repo; not the same as Cursor plugins

Why this topic matters now

  • Engineers work in client repos, scratch folders, and other org repos; the richest skills in brainforge-platform do not load there.
  • The team marketplace is the intended org-wide channel for “install once, use everywhere” plugin-shaped bundles (Cursor plugins reference).
  • Risk: Duplicating large trees into brainforge-plugin-marketplace creates drift and doubles review load unless there is a single source of truth or automated publish.

Past learnings (from repo docs)

  • A dedicated marketplace repo is already the right shell; the monorepo should not be the import URL for the team marketplace (see plan above).
  • Submodule-backed skills (Anthropic packs, Impeccable, Omni, last30days, etc.) are bootstrapped by npm in the monorepo — those are not “Brainforge-authored” and should likely stay upstream or marketplace vendor plugins, not copied into a “Brainforge bundle” wholesale.

2. Raw candidates (six frames → merged)

Ideas generated across pain / friction, inversion or automation, reframing, leverage, cross-domain analogy, constraint flip; near-duplicates merged.

A. Packaging and source of truth

  1. Single brainforge-standards marketplace plugin — all org-authored skills + rules in one plugin folder; monorepo deletes copies and links to marketplace only.
  2. Split plugins — e.g. brainforge-delivery, brainforge-gtm, brainforge-platform-dev as separate marketplace entries to limit noise and requiredness scope.
  3. Monorepo remains SoT; CI publishes tarball into marketplace repo on tag (no hand-editing skills in two places).
  4. Marketplace repo is SoT; monorepo submodules the plugin ( inversion: app repo consumes the published bundle).
  5. “Thin marketplace, fat docs” — marketplace ships only rules/ + pointers; long playbooks stay in a public or internal doc site with stable URLs in skills.

B. What moves vs what stays

  1. Move only alwaysApply: false / on-demand skills; keep security- and repo-specific rules in brainforge-platform only.
  2. Move only rules; keep skills in monorepo (smaller bundle, less duplication ofSubmodule trees).
  3. Move skills; keep rules in monorepo (skills are what people miss cross-repo; rules are often path-globs to this codebase).
  4. Extract a “Brainforge core” of ~10 skills (ship loop, Linear bridge, mission-run, keep-running pointers) and leave the long tail in the monorepo.
  5. Nothing moves — instead improve user-level sync (scheduled job to pull from a release artifact into ~/.cursor / user skills). (Constraint flip: marketplace is optional.)

C. MCP and secrets

  1. MCP templates only in marketplacemcp.json with ${VAR} placeholders; same file shape as today’s vendor plugins.
  2. No MCP in org bundle — every engineer already gets vendor plugins from the marketplace; Brainforge only ships skills/rules.
  3. “MCP doctor” skill in marketplace — one skill that encodes how to log in to Slack/Linear/Notion per AGENTS, without storing tokens.

D. Submodule and third-party skills

  1. Re-list submodule skills as separate marketplace dependencies (Omni, Impeccable, etc.) — do not vendor into brainforge-standards.
  2. One meta-skill that says “enable these plugins from the team marketplace in this order.”
  3. Pin submodule versions in marketplace via copy at release — reject as high drift (already rejected mentally; list for critique).

E. Codex / OpenCode parity

  1. Document-only in AGENTS.md: “After marketplace install, run X script to sync to Codex.” No dual publish.
  2. Parallel publish job — same content rendered to .cursor-plugin and a Codex pack zip; second-class but automated.
  3. OpenCode out of scope for v1; revisit when OpenCode has a team channel comparable to Cursor Teams.

F. Governance and ergonomics

  1. CODEOWNERS on marketplace repo for plugins/brainforge-*/ only Platform + one designee.
  2. Version line in plugin.json must match a changelog in monorepo knowledge/ or docs/.
  3. “Required” only for a minimal plugin (e.g. brainforge-core & Compound); heavy vertical skills stay optional.
  4. Quarterly prune — survivorship review like any internal package; archive skills nobody invokes.

3. Critique — what to reject or narrow

IdeaVerdictWhy
1 — one mega plugin with everythingReject as statedBundle size, cognitive load, and “required for all” becomes heavy; blurs GTM vs eng vs platform; Submodule/third-party content must not be copy-pasted in. Narrow: one plugin with curated Brainforge-only assets only.
4 — marketplace as sole SoT with monorepo as consumerReject for nowInverts how engineers edit skills today (they live and PR in the monorepo); high migration cost; conflicts with AGENTS.md and submodule story. Revisit only if the org commits to a packages-style skill repo.
7 — rules only, skills stayWeakCross-workspace value is skills; rules without globs to this repo help less than skills. Demote unless legal/compliance says “rules must not leave VPN boundary.”
10 — nothing in marketplace, only user syncReject as org defaultDoes not use Required / team distribution; depends on every engineer running scripts; no central version pin. Viable for opt-in power users only.
12 — no MCP in Brainforge bundlePartially rejectStubs and ${VAR} templates in marketplace are still useful for discoverability; no secrets in repo. Keep thin MCP templates optional.
16 — copy submodule trees into marketplaceRejectDrift from upstream; use vendor plugins in marketplace.json instead.
18 — full Codex zip parity in v1DeferDoubles build pipeline; get Cursor path working and adopted first.

4. Survivors (ranked)

RankIdeaTypeEffortWhy it survives
1metadata.pluginRoot + thin brainforge-standards (or rename) plugin with org-authored skills/rules only; vendor/Submodule skills remain separate marketplace entries (14)ArchitectureMedMatches “portable across workspaces” without forking Every/Omni/etc.; aligns with existing plugins/brainforge-platform slot (expand with real content or rename for clarity).
2Monorepo SoT + CI publish to brainforge-plugin-marketplace on tag or main (3)ProcessMed–HighOne authoritative edit location; marketplace gets reproducible updates; reviews stay in one PR flow if publish is automated.
3Split optional plugins by concern (2) after v1 if one bundle is too big — e.g. brainforge-delivery vs brainforge-gtmScalingMedUnblocks “required = small” (22) while keeping optional depth for specialists.
4Survivor 9 — start with a tight “core” list (ship loop, Linear/PR skills, mission-run/keep-running pointers, internal playbook links)Scope cutLowProves value before migrating dozens of skills; reduces first migration risk.
5MCP: templates + skill 11/13 — placeholder mcp.json or env-only patterns + one “connect your MCPs” skill linking to AGENTS sectionsSecurity-aware DXLowImproves onboarding without putting secrets in git.
65 — “thin marketplace, fat docs” for long playbooksHybridLowSkills stay short; knowledge/ and Notion stay canonical for walls of text.
720–22 — governance (CODEOWNERS, minimal “required” set)PeopleLowPrevents marketplace sprawl and policy debates later.
817 — post-install Codex mirror doc pathDocsLowRespects current sync-cursor-skills reality without blocking Cursor marketplace launch.

5. Synthesis — not a build plan (handoff to ce:brainstorm / ce:plan)

Recommended directional sequence (to be turned into a real plan with owners and acceptance tests):

  1. Define the inventory — classify every .cursor/skills/* and critical .cursor/rules/* as: (a) org-generic portable, (b) brainforge-platform path- or app-specific, (c) third-party / submodule. Only (a) + a short list of (b) reframed to be path-agnostic are marketplace candidates.
  2. Name the home plugin — either grow plugins/brainforge-platform into the real bundle or add plugins/brainforge-standards and deprecate the empty name; document the difference from vendor plugins in brainforge-plugin-marketplace/README.md.
  3. Choose SoT — default monorepo SoT + automated publish (Survivor 2) unless leadership wants a skills-only repo later.
  4. Cut scope v1 to Survivor 4 core set + governance (Survivors 7).
  5. Codex — keep documented mirror for v1 (Survivor 8); defer full dual-target packaging (rejected 18) until adoption metrics say so.
  6. After v1 works: consider plugin split (Survivor 3) and quarterly prune (23).

6. Post-ideation menu (per ce:ideate handoff)

  • Refine — narrow the “core 10” list with Platform + one GTM and one Delivery owner.
  • Open in Proof — if you use Proof for stakeholder review.
  • /ce:brainstorm — one chosen direction (e.g. “monorepo SoT + CI publish to plugins/brainforge-standards only”) to define constraints and non-goals precisely.
  • /ce:plan — implementation: CI shape, folder layout, marketplace.json entries, and AGENTS.md updates.
  • Save and end — this file is the save for ideation; no requirements or code produced here.