What Magicdoor Auto Shares With Jev (and What It Does Not)

Direct answer: For each Experimental Auto message, Magicdoor asks TypeSafe AI's Jev to evaluate a bounded effective text state, not the raw attached files or every full conversation turn. That state includes up to six effective user/assistant text messages, the latest request again, and a boolean capability flag. In a compacted chat, one of those messages can be a synthetic summary of older history, including artifact metadata and extracted descriptions. It can also include sensitive information typed into ordinary messages. The selected answer provider receives the context needed to answer through its own path; Jev only supplies routing signals.

Fact-checked October 2, 2026. This page owns the routing data-flow boundary, not what Jev is, Magicdoor's exact selection policy, or a provider-wide privacy certification.

Two separate paths

StepWhat happensImportant boundary
Before routingMagicdoor loads effective conversation history and checks required capabilitiesSaved file or container context can constrain the answer route even when the latest turn is text-only; compacted history can prepend a synthetic summary
Jev routing callA server request sends the bounded state and two typed questions to TypeSafe AI's jev-latestJev returns probabilities for substantive judgment and code/data execution; it does not draft the answer
Answer generationMagicdoor applies its quality-first-v7 policy and calls a concrete answer modelThe answer model receives the applicable conversation and attachments by the existing answer path; this is not the Jev state
Usage and storageMagicdoor records usage for concrete answer-model attempts and handles chat persistenceJev routing usage is operational telemetry, not added to answer-model usage; chat storage is separate from Jev

TypeSafe's State documentation explains that state is the content submitted for evaluation alongside questions. What is in Magicdoor's state is set by Magicdoor's server code, not by TypeSafe's generic examples.

Included and excluded at the Jev boundary

  • Included: text from at most six effective user/assistant messages (including the latest user turn), the latest request as a separate field, and requiresOpenAI as a boolean. In compacted saved chats with at most five newer messages, the synthetic user-role summary can be among those six. It may contain earlier conversation summaries and an artifact manifest with filenames, storage paths, IDs, and extracted descriptions. The JSON-serialized state is capped at 16,000 UTF-8 bytes. If the last eligible message has no usable text, or the serialized state exceeds that cap, the router takes a capability-aware fallback rather than chopping off text to force a cheaper route.
  • Excluded as raw parts: attached files and their URLs, tool payloads, system messages, separate reasoning parts, and embedded <think> blocks. This does not exclude attachment-derived details already present in a compacted text summary or text pasted by a user. The file-part rule concerns the Jev call, not whether an answer model can use a file. The flag can reveal a capability requirement even without sending its underlying attachment.
  • Not redacted in text: a user can paste an email address, URL, secret or file contents into an ordinary message. Those characters remain text eligible for the bounded state. Do not assume this is full redaction or a zero-data-retention promise.

For example, on a text-only follow-up after a PDF, Jev does not receive the PDF file part; in a compacted chat its synthetic text summary may describe the PDF or include its filename or storage path, and effective history can still require an OpenAI-capable answer model. If you type a customer identifier directly in your question, it may be sent as text to Jev. Choose a manual model to bypass Jev routing if that boundary is unsuitable; manual selection does not remove the selected answer provider or Magicdoor's storage path.

Limits you should understand

A timeout (2.5 seconds), service error, invalid result, missing or oversized state fails closed to a compatible answer model (normally Claude Opus 5.5 for compatible text, GPT-6.1 Sol when OpenAI capability is required). Cancellation does not launch another model. Auto may select DeepSeek V4.1 Flash, GPT-6 Luna, GPT-6.1 Sol, or Claude Opus 5.5, with an early DeepSeek-to-Opus handoff possible. This four-model pool is not a promise of the best answer. Read how to evaluate a router and what Magicdoor actually tested before interpreting the evidence.

Try Experimental Auto in chat

FAQ

Does Magicdoor send my files to Jev to choose a model?

Raw file parts and their attached URLs are excluded as parts, but pasted text and a compacted-history summary can include file-derived descriptions, filenames, IDs or storage paths. A capability flag can reflect an OpenAI requirement. The answer provider remains a separate boundary.

Does Auto redact all sensitive information before Jev sees it?

No. Raw file, tool, system and separate reasoning parts are excluded, but message text and compacted-history summaries can contain names, sensitive strings, URLs or artifact metadata. Exclusion is not general redaction.

Does Jev store or answer the full conversation?

Jev receives a bounded effective text state that can include a summary of older compacted history; it is not the answer model. This page does not assert TypeSafe retention or security guarantees. Magicdoor's chat storage and the selected answer provider are separate boundaries.

Sources

Accessed October 2, 2026.

  • TypeSafe AI: State, for the distinction between state, questions and model output.
  • Magicdoor's current typesafeRouter.ts, autoRouter.ts, routing policy, model registry and server answer flow for application-specific boundaries.

Copyright © 2026 magicdoor.ai