haft

Roadmap

Haft is actively developed. This page separates shipped v8 work from research directions; treat the latter as direction, not a promise.

Recently shipped (v8 / v8.1)

  • Governance substrate pivot — skills + CLI + MCP replaced the standalone agent, TUI, and desktop wrappers.
  • 15-skill catalog — haft init installs the FPF workflow skills for Claude Code and Codex.
  • Fused code graph — code symbols, callers, callees, impact, source, and governing decisions surface through haft_query.
  • Spec-first onboarding — active TargetSystemSpec, EnablingSystemSpec, and TermMap sections define project readiness.
  • Hybrid FPF search — keyword search fused with local semantic recall over baked FPF section vectors.
  • Shared embedding runtime — v8.1 reuses one local haft-embed sidecar across live haft serve sessions.
  • CI governance gate — haft check reports stale, drifted, unassessed, and coverage-gap findings.

Shared team state

A future shared server would manage the decision database for teams. Solo use stays local-first: MCP clients talk to haft serve, and haft check remains the CI gate. 

Solo:  Agent → MCP → haft serve → local SQLite
Team:  Agent → MCP → haft serve → haft server (remote) → PostgreSQL

A server can later add real-time collaboration and shared search. It does not replace the local artifact graph or the deterministic check command.

Formal specification (research)

Decision records from the full cycle contain invariants, post-conditions, and acceptance criteria — enough structure for any agent to generate property-based tests today. This works without any haft changes.

The research direction: can we go further? Translating decision invariants into Quint specifications (TLA+ with TypeScript syntax) for model checking — verifying that invariants hold across all reachable states, not just random inputs. This is in research phase.

Cognitive review loops

Instead of dumb valid_until timers, use FSRS (Free Spaced Repetition Scheduler) to schedule decision reviews based on evidence confidence and historical review quality. Requires the server mode (long-running process).

Autonomy budgets

From FPF E.16: treat agent autonomy as a managed resource (like money or risk), not as "magic of a smart model." Typed budgets with guards, separation of duties, override protocols. This is the team governance layer.

What's not on the roadmap

  • AI model lock-in — haft works with any MCP-compatible AI tool. We don't pick your model.
  • A replacement coding agent — v8 plugs into Claude Code, Codex, and other MCP hosts instead of shipping its own agent runtime.
  • Prescriptive methodology — haft provides structure, not opinions. Use as much or as little of the FPF cycle as fits your workflow.

Want to contribute?

Check the issues on GitHub. The contributing guide explains the workflow: issue first, then PR to dev.