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.

Provider protocols and configuration

provider_api selects the message protocol. The Python reference runtime supports Chat Completions, Responses, and native Anthropic. Compilation needs no model credentials and makes no provider request. Importing compiler or runtime modules may initialize configuration objects and read local .nxs files. Start with First provider setup.

Value	Selection and credentials	Preserved session content
chat	Default; existing .nxs and provider configuration	Chat messages, tool calls and results
responses	Explicit; existing OpenAI client configuration	Raw output items, encrypted reasoning, native tool results
anthropic	Explicit, or the anthropic/ model prefix	Raw content blocks, thinking signatures, native tool_result blocks

Declare the protocol

This complete declaration module compiles offline. Execution also requires an available model and credentials.

agent Analyst {
    prompt: "Analyze the supplied evidence."
    provider_api: "responses"
    timeout: 120
    max_tool_calls: 8
}

Set model in the declaration or use the effective configuration. use config name; selects an existing configuration; see Configuration. Supply credentials through the local environment or existing .nxs configuration, keeping them out of .nx files and version control.

For native Anthropic, install the extra in your Nexa checkout and set ANTHROPIC_API_KEY in the execution environment. ANTHROPIC_BASE_URL optionally selects the endpoint.

uv sync --locked --extra anthropic

Completion and failure

Streaming, non-streaming, Protocol output, and cache hits share output validation and completion handling. Truncation, refusal, or a stream without a completion event fails the run. Opaque provider items needed for continuation are checkpointed and restored with their original protocol.

Exact response cache

Caching applies to fresh Chat runs without tools. The key includes effective context, system instructions, model, endpoint, and output contract. Context limits are checked before a hit is accepted. Responses and Anthropic responses are currently uncached. Manage conversation history, provider state, and caches as separate resources.

Cache effectiveness depends on the workload. Offline fixtures check key isolation, output validation, and skipped requests. Measure model quality, cost, and latency separately with a selected provider and budget.

Continue with Execution and budgets and Context structure.

---

Source and validation · providers / cache