Replace the Vercel AI SDK with direct OpenAI SDK calls and a custom
AgentLoop. The AI SDK has zero runtime integration today, so removing
it costs nothing. Supply chain risk (2-5 releases/day, April 2026
Vercel breach, bus factor of 1) makes it a liability we don't need.
Key changes:
- ADR-018 accepted: openai package (zero runtime deps) replaces ai SDK
- AgentLoop handles multi-step tool execution explicitly (~300 LOC vs
AI SDK's ~2700 LOC streamText)
- Hub owns UIMessage/UIPart/ToolCallState types (extends ADR-016)
- Hub owns streaming protocol (subset of AI SDK's UIMessageChunk wire
format with step boundaries, error handling, usage tracking)
- operationToOpenAITool() maps TypeBox schemas directly, no adapter
- Trade-off: ~1100 LOC total new code for the savings of 6+ transitive
deps, supply chain risk, and release cadence coupling
Updates AGENTS.md constraints and dependencies, adds OQ-63/OQ-64/OQ-65
and Theme 11 (Inference & LLM Integration) to open questions.
ADR-014 (Docker-first deployment): resolves OQ-21, OQ-22, OQ-23,
OQ-34, OQ-35, OQ-36, OQ-37. Docker is the primary deployment model.
Redis/Postgres in same network. Config via mounted volumes. Single-
container restart for v1. Migrations block startup.
ADR-015 (Dev spoke, not opencode): resolves OQ-16, OQ-17, OQ-26,
OQ-28, OQ-51, OQ-55. Replaces opencode integration with a compiled
dev spoke binary. Hub owns session format. Opencode compat is an
import tool, not an architectural constraint. Adds OQ-61 (dev spoke
operations) and OQ-62 (dev spoke distribution).
ADR-016 (Hub-own schema): resolves OQ-18, OQ-19. Hub defines its own
canonical message/part format. JSONB is implicitly versioned. Flat
parts for v1. Compaction is a hub concern (pruning), not opencode's.
ADR-017 (Hub-first roles): resolves OQ-26, OQ-28, OQ-51 (overlapping
with ADR-015). Hub is database-first for roles. Seeded by migrations.
No file sync needed. hub.createRole for custom roles.
Also narrowed: OQ-04 (service accounts), OQ-05 (git SSO out of scope),
OQ-08 (spoke-side concern), OQ-09 (v1: reconnect only), OQ-11 (dev
spoke replaces container spoke), OQ-29 (hub-only concern), OQ-41
(gitea ops are optional spoke concern).
Deferred: OQ-52 (memory), OQ-55 (anthropic import).
Net result: 15 resolved, 7 narrowed, 2 deferred out of 62 total.
39 remain open, down from 60 in the original tracker.
60 open questions across 10 themes, synthesized from 20+ architecture
and storage documents. Follows the flowgraph open-questions.md pattern
with ID tracking, cross-references, dependency chains, and a priority
resolution order for the 11 high-priority blocking questions.
