The substrate is files
Solo mode is YAML on disk: one file per session record, one per task, under dated directories. No database required to participate, nothing to host, everything reviewable in a diff. The ontology and event vocabularies are canonical YAML files with TypeScript Zod mirrors, and a CI check fails the build if the two ever disagree.
The skills
The working surface is Claude Code itself: /clear-review and /session-note slash commands, plus a doc-system kit (a /doc-sync orchestrator and doc-steward / doc-reviewer sub-agents) and a markdown doc linter. Three further skills are specified but unbuilt, waiting until the corpus is big enough to justify them.
Team mode
The multi-user product layers on top without replacing the file substrate:
- Web: Next.js 15 dashboard on Vercel for tasks, sessions, drift signals, and predictions.
- MCP server: Hono with HTTP + SSE on Railway, exposing 23 tools for substrate reads and writes, every one of which emits to an append-only audit ledger (23 event types).
- CLI:
retroctl, for the same operations from a terminal. - Database: Supabase Postgres via Drizzle, 15 migrations, row-level security, with session detail held as JSONB event captures rather than flattened columns.
Verification layers
Tier 1 is the drafting agent auditing itself with required evidence references. Tier 2 dispatches a fresh, smaller-model agent to independently verify high-consequence entries, and disagreement between the tiers is preserved rather than overwritten. The same pluggable library backs drift detection.