Skip to content

Roadmap

reactifact is pre-1.0, one maintainer. This page is the honest current state of "what's next" — not a wishlist. See CHANGELOG.md for what's already shipped, release by release.

Now

  • Finish the Step-era API. The reactive core is being consolidated around Step/@step, step-owned consumes/produces, explicit effects: Effects, typed inputs (Artifact[T], list[Artifact[T]], Annotated[T, When(...)], Latest[T]) and Agent.steps. The goal is one obvious way to write a unit of work.
  • Content-addressed artifact state. Artifact is a frozen handle over an append-only ArtifactStore; Context.artifact_version / version_history / find_versions are canonical, persistence is v2 (blobs + version index + commit data_ref), and compaction is authoritative. Remaining work is closing compatibility edges and docs.
  • Keep the example set small. The repo keeps the reference applications (devops, knowledge, forklab, repair, medic_lab, fintech_audit, starter_app, chat) instead of maintaining every historical pattern demo.
  • MCP hardening (more transport coverage, more real-server testing) before adding new integration surfaces.

Next

  • Stable 1.0 — freeze the public API surface, commit to semver guarantees, and close out breaking changes before they accumulate further. The surface has three stability tiers:
  • Core (frozen hardest): Context/Artifact (frozen handle + version store)/Effects/Patch/Step/Agent/Source/Provenance/HITL (PendingQuestion, effects.ask/resume), plus RuntimeResources's own extension points — context_builder (reactifact.context_builder) and verification_threshold (reactifact.verify) count as core because they're config surface on a core object, not a separable add-on.
  • "In the box" (stable, evolves faster than core): Tool/ToolUse/ ToolUseHITL (including the destructive-tool approval gate), reactifact.tools.native, reactifact.tools.agent.AgentAsTool, verify.Verify's metric surface, MCP, tracing/observability, providers. Same semver discipline, a lower bar for additive capability.

  • Recipes (reactifact.recipes, docs/en/recipes.md) — frozen at 1.0, explicitly the layer meant to keep growing fastest post-1.0; new recipes are additive, framework-internal moves stay off their public surface.

  • More Source integrations — today's built-in sources are filesystem, CSV, embeddings and WebSource (see the knowledge example). The Source protocol (reactifact/sources.py) is intentionally small so this grows by adding implementations, not by changing the abstraction: direct API sources, SQL and keyword search are the concrete gaps.

Non-goals (for now)

Carried over from docs/en/comparison.md — worth repeating here since a roadmap page is where people look for the opposite promise:

  • No managed hosting / SaaS platform. reactifact is a library; there is no hosted execution or UI for non-engineers planned.
  • No large pre-built agent/tool marketplace. The Source and MCP abstractions stay small and composable rather than growing a plugin ecosystem to compete on integration count.

Contributing to the direction

Roadmap changes and feature discussion happen in GitHub Discussions, not silently in code. If something here looks wrong or missing, open one.