Sparra started the way the workshop suggested: get the whole loop standing before polishing any one part of it. In one long day I wired the full spine end to end. Filesystem state that detects whether a project is greenfield or existing, a session layer on top of the Claude Agent SDK with tracing and guard hooks, the planning phases that co-edit a PLAN.md, the autonomous build loop with negotiated contracts and an adversarial evaluator, a reflect step that reads the run's traces, and a sparra CLI to drive it.
The test that the skeleton was real came almost immediately: I pointed Sparra at its own codebase and had it build parts of itself, using the vitest suite it generated as the regression gate. A harness that can work on itself is a harness I can actually trust to tell me when it is broken. Everything since has been filling in this frame rather than redrawing it.