Flow SpecKit

Durable, artifact-driven AI-SDLC workflow orchestration — think "Temporal for AI Software Engineering".

Flow SpecKit orchestrates the complete AI-assisted software development lifecycle, not just the coding step:

Business Idea → Discovery → Product Shaping → Technical Design
             → Implementation → Review → Merge → Release

Durable, resumable workflows drive stateless AI skills and pluggable coding agents (Claude Code, Cursor, …), with human approval gates owning every consequential transition and every phase producing a versioned, lineage-linked artifact.

Status: early development. Phase 2 (Artifact Engine) is complete; Phase 3 (Workflow Engine) is in progress. The full wedge demo below is the v0.1 target, not yet functional end to end.

Install

pip install flow-speckit

The wedge (v0.1 target demo)

In any existing repository, with nothing installed but Python and a coding-agent CLI:

$ flow-speckit init
$ flow-speckit run feature --idea "Add CSV export to the reports page"

which durably executes:

FrameBrief → [human gate: approve brief] → ProductArtifact → TechnicalDesign
          → [human gate: approve design] → Claude Code implements in an isolated
            git worktree → Pull request opened

Three properties no prompt-first tool can offer:

1. Durability — kill -9 the process mid-implementation, then flow-speckit resume continues exactly where it stopped.
2. Lineage — flow-speckit trace <pr-url> prints the full provenance of the PR back to the business idea: every artifact version, who approved what and when, and per-step token cost.
3. Accountability — the append-only event log proves the workflow, not the model, decided every transition; humans resolved every gate.

Core principles

• Artifact-driven — every phase consumes an artifact and produces a new, versioned, schema-validated artifact. Artifacts, never chat history, are the source of truth.
• Workflow-first — a durable state machine owns all transitions. AI never picks the next step.
• Stateless skills — every AI capability is a stateless, independently executable skill. No long-running autonomous agents.
• Human gates — approval gates are first-class workflow states. Humans own decisions; AI owns artifacts.
• Pluggable execution — coding work delegates to interchangeable backends (Claude Code first) behind a small adapter port.

Architecture at a glance

• Python 3.11+, single package — no Docker, no service stack.
• PostgreSQL-only core (embedded server via the embedded-pg extra): the work queue, event log, artifact graph, blobs, and full-text search all live in one database.
• Four kernel subsystems: Artifact Engine, Workflow Engine, Skill Engine, Execution Port. Everything else is a plugin behind an entry point.

Full design documents live in the repository: github.com/vinit-devops/flow-speckit.

License

Apache-2.0.