Decision Swap: codemod and router that moves classify-style LLM calls to decision models

The category is only weeks old. Jev arrived on September 15, OpenAI's Decisions API on October 6 and Microsoft-Decision-1 on October 8-9, 2026, and every tool so far works with one vendor. The migration CLI idea is partly taken by should-i-jev and prompt2jev, but both target only Jev. The single-endpoint idea is weak: OpenRouter's Decisions endpoint and Vercel AI Gateway already give one key across decision models, and existing routers can easily add a cost-or-quality rule. The clearest opening is a tool that works across all the vendors. It would find decision-shaped calls in Python and TypeScript, copy real traffic to Jev, OpenAI Decisions, Decision-1, Laya and others, and report agreement with the current LLM, calibration, sensitivity to option order and cost per thousand decisions. It would then write the adapter and roll it out gradually with a fallback to the LLM. Demand for this is real because the vendors' claims are unverified: Microsoft published a method but no benchmark results, and independent papers show Jev and Laya changing answers when options are reordered and giving poorly calibrated confidence. The selling point is evidence a team can trust before switching vendors, plus ongoing checks for drift. The main risks are that OpenRouter, Vercel or OpenAI ship their own migration and eval tools, that the shadow-testing piece can be rebuilt with existing eval tools, and that the market may stay small if vendors' prices drop to the point that savings are trivial. Clef-flash and H2O-4B turned up in no search, so confirm those models exist before building around them.

• should-i-jev (proposed in awesome-jev-tools PR #11)

  Described as a dependency-free CLI that scans LLM logs and code for decision-shaped calls, prices the switch to Jev and opens a migration pull request. Almost the same as the decision-swap CLI, but Jev-only. I couldn't find its repository, so whether it really exists and is maintained is unconfirmed.

• prompt2jev / llm-to-jev (listed in awesome-jev-tools)

  Community tools that turn LLM prompts, or the code that runs them, into Jev typed questions. They cover the adapter-writing step for a single vendor.

• OpenRouter Decisions endpoint and Jev prompt-to-questions lab

  Alpha endpoint (POST /api/alpha/decisions) with a live list of decision models, defaulting to typesafe/jev-1.13. One key for decision models, so it directly threatens the single-endpoint idea. A companion lab page rewrites prompts as Jev questions: https://openrouter.ai/labs/jev/compile

• OpenAI Decisions API

  Public beta since October 6, 2026 on gpt-6-luna, with predicate, choice and score question types, at $0.10 per million input tokens. OpenAI could easily ship its own migration help for its own SDK users.

• Microsoft-Decision-1 (Foundry, Vercel AI Gateway)

  Hosted scoring model launched October 8-9, 2026 at about $0.04 per million input tokens. Vercel AI Gateway already serves it alongside LLMs, so a gateway that holds both decision models and LLMs already exists.

• TypeSafe Jev

  The original decision-model API, with choice, score and yes/no question types. A target vendor, and it also ships an open-source adapter that makes ordinary LLMs return Jev-style decisions.

• Laya (Convai)

  Open-source (Apache 2.0) 421M decision model. Its repo includes a router that picks the right checkpoint per request. A cheap self-hosted option that a vendor-neutral tool would need to support.

• vLLM-SR Decision 1.0

  Open decision foundation models from the vLLM semantic-router project, which already routes requests, so a cost-based router could appear there.

• OrcaRouter

  LLM router company writing many comparison posts on decision models (Decision-1, Laya, Kev, Intern-Decision-4B). It says it doesn't host them yet, but it is clearly watching the category and could add a decision route.

• rightmodeler

  MIT-licensed command-line tool that replays production traces from Langfuse, Helicone or OpenTelemetry against cheaper models. Closest existing tool to the shadow-judge idea, but it only tests chat models, not decision APIs.

• burn0

  Command-line tool that scans a codebase for OpenAI, Anthropic and other API calls and tracks cost. It covers the code-scanning step but doesn't suggest switching to decision models.

• RouteLLM / Not Diamond / Martian

  Established routers that pick between a cheap and an expensive LLM by cost and quality. They don't target decision APIs today, but adding them would be a short step.

Checked 2026-10-10