What problem does it solve? Multi-agent builds often pass every unit test and static gate while the actual product is broken, because tests are written by the same agent that wrote the bug and no gate ever runs the artifact end to end. ## Core Features & Use Cases - Graph-Based Build Planning: Structures a build as a planning thread, parallel module workers with upfront goal contracts, and an ordered evaluation chain ending in packaging. - Runtime Evaluation Harness: Ships runtime-eval.sh, a bash harness that builds, serves, and walks a project's real user journey, failing loudly with the named broken step. - Self-Verified Gate: The harness includes a --self-test mode proving it reports failures as failures, so the gate cannot only ever pass. - Use Case: When a build is large enough to need multiple workers or must run unattended, write per-module goal contracts, spawn workers with only their contract, then run the runtime evaluation before any code review. ## Quick Start Run bash .claude/skills/agentic-coding-graph/runtime-eval.sh with your project name to execute its full user journey and see which steps pass or fail.
