2.4 compatibility in Nexa 3.0

This page describes the retained 2.4 language/runtime path. Sources without a version declaration use this path. Start new Open Programs with Nexa 3.0; compare contracts in the migration guide.

Structured context

An Agent's context declaration defines context source, sink, schemas, and inherited fields. Pipelines whose endpoints both declare context transfer AgentContext. Python applications can also call agent.run(previous.last_context()) explicitly.

Complete declarations

This module declares upstream and downstream Agents and compiles without a model.

protocol Findings { summary: string }
agent Researcher {
    prompt: "Extract the facts."
    context {
        source: upstream,
        sink: downstream,
        output_schema: Findings,
        inherit: [messages]
    }
}
agent Writer {
    prompt: "Write from the facts."
    context {
        source: upstream,
        sink: downstream,
        input_schema: Findings,
        inherit: [messages, artifacts]
    }
}

Field	Purpose
source / sink	Select upstream input and downstream output behavior
input_schema / output_schema	Declare context contracts, compared by nominal name equality
inherit	Select inherited message or artifact fields
max_history_turns	Bound retained conversation turns

shared:name shares deep-copied snapshots within one flow; discard creates no downstream context. Different schema names require explicit conversion even if their fields resemble each other.

Context budget

Context limits currently use serialized UTF-8 byte size as a conservative token upper bound. Provider usage is accounted separately. Chat trimming keeps complete user turns and preserves tool-call/result pairs. A newest turn that exceeds the limit fails.

Native Responses/Anthropic retain opaque items. When their history or context exceeds a limit, shorten input to preserve protocol state. See Provider protocols.

---

Source and validation · context / providers