Skip to main content

Roadmap

This is a living document. Priorities may shift based on Wave 1–3 learnings and community feedback.

Shipped (Wave 1)

  • Core block system — All four block kinds (handler, generator, sequencer, router)
  • SSE streaming — Sequence-number resume, item/content model
  • 4-scope state model — Request, session, user, project with resources, template rendering, and contentFile support
  • Eval framework — evalBlock, evalFlow, analyzerScorer (LLM-as-judge)
  • CLI harness — fsdev block, fsdev run with streaming NDJSON output
  • DevTool inspector app — Real-time flow visualization, trace view, session state panels, replay controls
  • npm publishing — Changesets + CI
  • Examples — Kitchen Sink and Hello Chat
  • External agent SDK integration research — Claude Code SDK wrapping

In Progress (Wave 1)

  • Execution trace system — Full block lifecycle event depth for the DevTool trace view
  • Transient block output — Stream-only items that bypass session persistence
  • Composable patterns library — Observer, Chain-of-Agents, and other reusable sequencer compositions
  • fsdev eval CLI command — Run eval suites from the terminal
  • fsdev dev — Launch the DevTool against your own flows with a single command
  • Session retention policies — Bounded item log with maxItems/maxAge eviction

Planned (Wave 2)

  • CLI TUI mode — Ink-based pretty-print rendering for human-friendly terminal output
  • Middleware extension system — Public interceptor API for caching, observability, and policy
  • Observability middleware package — @flow-state-dev/middleware-observability
  • Production eval middleware — Sample live traffic, score async, export to observability platforms

Planned (Wave 3)

  • Durable execution — Suspend/resume, checkpoint management, human-in-the-loop patterns
  • Production hardening — Request size limits, stream caps, security headers
  • Embedded in-app DevTools widget — First-party UI component for debugging flows in production

Research (Wave 4)

  • To be determined based on Wave 1–3 learnings and community feedback