Epic: BoJ capability-expansion roadmap (13 items)

[hyperpolymath]

Background

Strategic-direction issue capturing the 14 on-mission gaps surfaced in the 2026-05-20 session. Update 2026-05-20: Tier A code + all Tier B RFCs landed; Tier C remains.

BoJ's character: unified consolidation, formal verification, multi-agent coordination, trust tiers, security posture. Each item below was scored as "on-mission" rather than "could conceivably ship."

Tier A — implementable (real code)

• 1. MCP resources + prompts — bridge already declares resources + prompts capabilities in initialize; serve actual URIs (boj://cartridges/<name>/manifest, boj://capabilities/matrix, boj://proofs/<obligation>) and prompt templates ("audit my repo", "convene multi-agent cluster", "deploy with DNS+SSL"). ~300 LOC. Highest coherence-payoff. — DONE in PR feat(mcp-bridge): implement resources + prompts surface (epic #87 item 1) #89.
• 13. OTel structured traces — wire OTLP-compatible exporter into the bridge so every tool call emits a span. Pairs with observe-mcp / grafana-mcp in the cartridge set. ~150 LOC. — DONE in PR feat(mcp-bridge): OTLP/HTTP+JSON span emission (epic #87 item 13) #91 (zero-dep hand-rolled exporter; OFF by default, enabled via OTEL_EXPORTER_OTLP_ENDPOINT).
• 8. Search-engine cartridge — search-mcp covering Tavily / Brave / Exa / Perplexity behind one cartridge. Fundamental LLM-agent capability gap.
• 7. Vector database + RAG cartridges — pinecone-mcp, weaviate-mcp, qdrant-mcp, chromadb-mcp. Composes with local-memory-mcp + codeseeker-mcp for the full memory/RAG story.
• 9. Multi-modal cartridges — whisper-mcp (STT), elevenlabs-mcp (TTS), replicate-mcp (image/video), ffmpeg-mcp (transcoding). Glaring omission given multi-modal agent work.
• 14. Streamable HTTP transport — add MCP's HTTP+SSE transport alongside stdio so BoJ deploys as a Cloudflare Worker and browser-based clients can talk to a remote instance. Replaces stdio for remote clients without breaking stdio for local clients (transports coexist). ~60% of bridge tools (HTTP-based cartridges — GitHub, GitLab, Cloudflare, ML, research) work fully in Worker context; local-only cartridges (browser-mcp, container-mcp, local-coord-mcp) still require host-process deployment. Aligns with MCP roadmap direction; Cloudflare published a Workers+Durable Objects pattern. ~1-2 weeks focused work. NEW 2026-05-20 — added after JSR runtime-compatibility question surfaced the gap.

Tier B — RFCs (design-first)

• 4. Trust-tier policy DSL — extend Nickel beyond coord messages to per-cartridge-per-operation tier gating. PEP at the bridge; PDP as a separate policy-mcp cartridge. Turns Teranga/Shield/Ayo from concept into runtime enforcement. Aligns with ADR-0002 "BoJ-only MCP" (one gateway → one PEP). — DONE: ADR-0007 in PR rfc: ADR-0007 trust-tier policy DSL + ADR-0008 cartridge marketplace (epic #87 items 4 + 10) #92.
• 10. Cartridge marketplace — discovery protocol for Ayo-tier (community-submitted) cartridges. Currently no installation path for third-party cartridges; activates the third trust tier. — DONE: ADR-0008 in PR rfc: ADR-0007 trust-tier policy DSL + ADR-0008 cartridge marketplace (epic #87 items 4 + 10) #92.
• 5. Webhooks inbound + MCP notifications — HTTP listener architecture; inbound GitHub/Cloudflare/Sentry events → MCP notifications/*. Closes the agent feedback loop; agents become operators, not just queriers. — DONE: ADR-0011 in PR rfc: ADR-0011 webhooks inbound + ADR-0012 server-initiated sampling (epic #87 items 5 + 6) #95.
• 6. Sampling (server-initiated) — MCP sampling/createMessage for cartridge composition routing and ambiguous-input clarification. Underused capability across the MCP ecosystem. — DONE: ADR-0012 in PR rfc: ADR-0011 webhooks inbound + ADR-0012 server-initiated sampling (epic #87 items 5 + 6) #95.
• 2. Sandbox cartridge — sandbox-mcp wrapping e2b / Modal / CodeSandbox / Replit. Slots into the apprentice/journeyman/master tier model. Pairs with panic-attack-mcp (static pre-flight) → sandbox (dynamic execution) → vordr-mcp (post-execution integrity). — DONE: ADR-0009 in PR rfc: ADR-0009 sandbox cartridge + ADR-0010 cross-machine federation (epic #87 items 2 + 3) #93.
• 3. Cross-machine coord federation — DID-based peer identity, ML-KEM key exchange, federated quarantine. Loopback local-coord-mcp → encrypted mesh. Aligns with EXHIBIT-B quantum-safe provenance + ML-DSA-87 signing in stapeln.toml. The 0.1.0 Umoja gossip work is the starting point. — DONE: ADR-0010 in PR rfc: ADR-0009 sandbox cartridge + ADR-0010 cross-machine federation (epic #87 items 2 + 3) #93.

Tier C — multi-week campaigns

• 11. Discharge last 4 believe_me axioms — proof-debt audit reduced 31 → 4 in v0.4.0. Closing the last 4 lets BoJ claim "zero unproven axioms in the ABI" — rare in the wild.
• 12. Cross-cartridge composition proof — currently each cartridge is verified in isolation; cross-cartridge boj_cartridge_invoke is not proven-safe end-to-end. Research-grade work; JOSS-paper-shaped.

Scoring rationale

All 14 selected because they're on the BoJ character axis:

• Items 1, 13 → first-class MCP citizen
• Items 4, 10, 2, 3 → trust-tier + multi-agent coherence (the real ambition)
• Items 5, 6 → bidirectional agent loop (not just request/response)
• Items 7, 8, 9 → consolidation-mission completeness
• Item 14 → deployment-target breadth (Workers + remote browser clients)
• Items 11, 12 → formal-verification differentiation

What's NOT on this list (deliberate)

• Payments / crypto cartridges — outside BoJ's character; muddies trust model
• CRM cartridges — not on-mission; consolidation ≠ "everything"
• More git forges (Gitea, Bitbucket, Forgejo) — git-mcp generic covers them
• Yet-another-message-queue — queues-mcp already generic
• Making the existing stdio bridge run in a browser/Worker — architecturally wrong; item 14 is the right framing

Sequencing rule

Status 2026-05-20: Tier A items 1 + 13 in code; all 6 Tier B RFCs landed. Recommended next order:

1. Item 14 (Streamable HTTP transport) — write RFC first, then implement; biggest deployment-shape change
2. Item 8 (search-mcp) — fundamental gap, single cartridge
3. Items 7 / 9 — cartridge waves, can be parallelized
4. Items 11 / 12 — long-running proof campaigns, no rush