jev-codex-router vs DBOS

The two solve different problems at different layers. jev-codex-router is a routing layer: it makes a model-plus-thinking-effort decision for every model call rather than once per session, using four Choice questions per request (Astra policy, capability tier, effort depth, route lease) and a tier ladder of Luna → Terra → Sol → Astra with routing priors. Routing applies to continuations after tool calls, not just the first turn; responses pass through as Responses-in, Responses-out with the SSE stream relayed verbatim. Two projections run independently — Jev sees bounded decision state while the model gets full canonical replay. Reliability leans on fail-open behavior (a Jev error keeps the turn alive on a fallback route), a sentinel-file kill switch, and quota fallback triggered only on observed native quota exhaustion, retrying only when a distinct candidate exists. Every routed turn logs locally to ~/.codex/codex-router/jev-router-live.jsonl, never published.

DBOS is durable execution, not routing. Ordinary functions are annotated as workflows and steps in TypeScript, Python, Go, and Java — DBOS for Rust landed June 15, 2026 — so a crash, redeploy, or timeout replays from the last completed checkpoint instead of restarting. It adds durable queues with configurable concurrency and start rates, human-in-the-loop pause/resume via durable send/recv, durable sleep measured in days or weeks, dynamic cron schedules maintained from code, and Conductor for real-time workflow and queue monitoring. Recent news extends this: a September 17, 2026 architecture post on durable execution internals, an August 20, 2026 explainer on Conductor as a control plane, and Postgres-scaling benchmarks on SELECT DISTINCT and LISTEN/NOTIFY.

jev-codex-router is free as far as this data goes: an MIT monorepo on GitHub (0xNatoshi/jev-codex-router) with no license fee, no hosted tier, and no support contract listed. The real cost is your own: standing up and maintaining a local server component, working from a README, install.sh, and markdown docs. There is no published release-notes cadence, no SLA and no compliance certifications, and — importantly for a buyer's business case — no live measured quota-savings number you can take to a manager. It does embed a maintained Codex Router fork under router/ and connects Jev through that fork's generic-provider and curated-model extension points, so you avoid a second checkout, submodule, or hidden clone. Budget engineering hours, not subscription dollars.

DBOS is freemium: the open-source DBOS Transact package is the entry point and DBOS Cloud carries paid tiers. The data flags a specific commercial gap rather than a dollar figure — teams get stuck between Pro's 2 seats / 3 apps and the $499/mo jump to Teams. So the cost question splits: OSS is free but you run it, while the managed path forces an early seats-and-apps ceiling that makes mid-size teams decide whether the jump is worth it. As a bonus, DBOS doesn't charge for state storage because it never stores your data — workflow state stays in the Postgres you already pay for.