Function build_agent_node_handler 

pub async fn build_agent_node_handler(
    merged_agents: HashMap<String, AgentConfig>,
    tenant: String,
    secrets: Arc<dyn SecretsManager>,
    packs: Vec<Arc<PackRuntime>>,
    audit_sink: Option<AuditSink>,
    project_id: Option<String>,
) -> Option<Arc<dyn AgentNodeHandler>>

Build the production DwAgent handler if the environment is configured.

Returns None (so DwAgent flow dispatch errors clearly) under any of these graceful-degradation conditions:

• merged_agents is empty (no agents from packs or operator config);
• GREENTIC_AW_REDIS_URL is unset/empty;
• the AW Redis connection fails;
• the extension runtime fails to initialise.

merged_agents is the result of merging pack-embedded agents (base) with operator-declared [HostConfig::agents] (operator wins on collision). This merged map replaces the former direct read of config.agents so pack-provided agents are included in the runtime.

Redis is sourced from the environment because the runner uses an in-memory flow-state store by default and carries no Redis URL in [HostConfig]; this mirrors the existing env-config convention.

The tenant and secrets arguments wire in per-tenant LLM credential resolution when GREENTIC_AW_LLM_EXTENSION is set: requests resolve credentials from the secrets broker for the calling tenant rather than reading global env vars. Callers without per-tenant context (e.g. serve_agentic) should use build_agent_runtime directly, which uses the env-keyed backend and accepts no secrets context.

audit_sink (EPIC-B B-3) is threaded straight through to the constructed RuntimeAgentNodeHandler; None (no NATS audit client configured) keeps dw.agent execution on the plain greentic_aw_runtime::AgentRuntime::step path — zero behaviour change from before this parameter existed.

project_id is the deployed unit’s bundle_id — the value greentic-designer records as pack_name — used as the project_id billing dimension. Pass None when no bundle is pinned (the legacy tenant-only pack path); billing then omits the dimension instead of attributing spend to a non-unique agent id.