jev-codex-router vs Temporal AI

Temporal's differentiator is state: Workflow state is captured at every step with no checkpointing or recovery code, Activities retry with automatic backoff plus four timeout classes and heartbeating, and Durable Timers can sleep for months while cron Schedules support backfill and Continue-As-New. Execution is interactive — Signals, Queries, and Updates read and mutate a running Workflow mid-flight, and Workflow Streams add real-time interactivity. Multi-language is broad rather than deep: native SDKs for Go, Java, Python, TypeScript, .NET, PHP, Ruby, and Rust. Agent-specific tooling lands as Activities (OpenAI Agents SDK, Google ADK) alongside LangGraph and LlamaIndex integrations; Child Workflows handle fault isolation and Nexus handles durable cross-team calls. The trade-off is determinism — no random calls or direct clock reads in Workflow code.

jev-codex-router solves a different problem: instead of one global model setting for a Codex session, it decides model plus reasoning/thinking effort per call, and applies that to continuations after tool calls rather than only the opening turn. It asks four Choice questions per request (Astra policy, capability tier, effort depth, route lease) and walks an Luna → Terra → Sol → Astra tier ladder with routing priors. Astra policy is mandatory for architecture and risk-focused review. It is Responses-in, Responses-out with the SSE stream relayed verbatim, no format conversion, and two projections so the bounded decision state stays separate from the canonical replay. There is no durability, no scheduling, no state persistence — just routing, local logging, and a kill switch.

Temporal is freemium in the sense that the durable execution model and SDKs are open source and self-hostable, with Temporal Cloud as the paid managed path and an enterprise tier for teams that need audit logging, SSO, and SOC 2/HIPAA on an SLA-backed support path. Recent Cloud work has been about organization and reach rather than pricing changes: Pre-Release Projects organize namespaces and Nexus endpoints, Custom Roles add granular resource-level permissions (Pre-Release), and Cloud on Azure is invite-only Pre-Release. Serverless Workers matter here because they cut your infrastructure bill directly — AWS Lambda is in Public Preview with Temporal controlling worker scaling and lifecycle, and GCP Cloud Run is in pre-release with Temporal scaling Worker Pools in response to available work. The hidden cost is architectural: deterministic Workflow code and the Workflow/Activity split are a real learning curve, and the docs themselves say Airflow or Argo is simpler and cheaper for batch DAG pipelines.

jev-codex-router costs nothing — MIT-licensed, distributed as a self-contained monorepo (0xNatoshi/jev-codex-router) with an embedded Codex Router fork, so no second checkout or submodule. The router's economic story is quota: it aims to pick the least expensive configuration with enough capability for the next decision, a sufficiency objective rather than a maximum-intelligence one. Note that the vendor explicitly lists "users who need a live, measured quota-savings number