Status: accepted for specification; implementation pending.
How should graph state manipulate model context without passing the complete conversation, project graph, workspace, retrieval store, and tool history into every agent node?
Every model-executed TLG node runs in a fresh Flue child-session capsule over a sealed, versioned context envelope assembled by an application-owned context controller.
The envelope includes mandatory direct inputs, protocols, authority, approvals, safety constraints, accepted upstream evidence, declared capabilities, and budgeted optional context. It records inclusion and omission reasons, token accounting, provenance, and a digest.
The capsule has no undeclared generic filesystem, graph, transcript, retrieval, web, or credential access. Additional information or action requires a typed broker request and a newly sealed envelope.
Typed graph state alone does not bound an LLM prompt. Context control requires an enforced adapter boundary that determines the exact payload and capability surface for each invocation.
plan-implementationimplement-core-contractsimplement-agent-runtimeimplement-memory-retrievalimplement-capabilities-securityimplement-ingress-operationsintegrate-and-repairreview-architecture-securityspecify-task-lifecycle-architectureverify-release-reconciliation-specificationsRevisit only when another mechanism can prove the exact model payload, capability absence, provenance, mandatory closure, and invalidation behavior at least as strongly.