graph-engineering

Build an execution graph: fix a verification contract from the governing specs, cut nodes at judging seams with bounded write scopes, draw typed edges (needs, informs, excludes), and schedule writers serially by default. Render the graph as Mermaid and get user approval before dispatching any node.

Use a graph only when units depend on, inform, or exclude each other, such as shared write scopes or single-writer resources. When no unit writes and none depends on another, run plain concurrent reads with no contract, ledger, or verifier.

Yes. On Claude Code it maps nodes to agent() calls in a Workflow script with schema-enforced outputs and worktree isolation. On Codex it maps nodes to durable managed threads, with the coordinator maintaining a written ledger as the authoritative schedule.

Each node gets a fresh-context verifier that receives only the assertions and the real artifact, never the worker's narration or rationale. Verifiers run static checks and exercise the running system, returning pass, fail, or unjudgeable per assertion with recorded commands and exit codes.

A heartbeat monitor tracks headroom and triggers a pause protocol at 90% of any limit: the frontier freezes, in-flight nodes write partial handoffs, the ledger is saved, and a resume is scheduled for the limit window reset. On resume, the frontier is recomputed from actual repository state.

Concurrent writers are allowed only when all four conditions hold: disjoint owned paths including lockfiles, no shared open decision, independent verifiability, and workspace-level isolation such as separate worktrees. Otherwise writes stay serial because reconciliation costs more than the time saved.