weft

Weft

Spec-driven AI development orchestration, woven on jj and beads.

Status: design — see docs/design.md. No engine yet.

Weft is a meta-prompting / context-engineering system for AI coding agents. It takes the methodology proven by GSD — a tight, spec-driven loop that keeps the main context clean by doing heavy work in fresh subagent contexts — and rebuilds it on purpose-built substrates instead of homegrown ones:

• The warp — beads (bd) is the brain. It owns planning, the dependency graph, task state, and scheduling. bd ready is the scheduler. There is no ROADMAP.md / STATE.md / SUMMARY.md.
• The weft — execution. Agent work woven across the warp in parallel passes (waves), each pass laying down one change per ready bead.
• The loom — jj (Jujutsu) is the substrate. It holds the warp under tension, lets passes run in parallel sheds (isolated workspaces), and tolerates a dropped pick without unravelling the fabric (first-class conflicts).

The name says the design: beads holds the structure; jj does the weaving.

Why

GSD invented a file-based tracker (.planning/) and a git-commit choreography because it had no real tools underneath. beads is the purpose-built tracker; jj is the purpose-built VCS. Weft keeps GSD's methodology and deletes the two subsystems GSD only built to compensate.

Vocabulary

Term	Meaning
warp	the bead dependency graph — the plan, held under tension
weft	the woven work — agent changes laid across the warp
pick	a single woven change (one bead → one jj change)
shed	a parallel wave — the set of ready beads woven together

License

Apache License 2.0. See LICENSE.

Methodology and several command/agent prompts are adapted from GSD Core (MIT); see docs/design.md for attribution.