Auto-Router Updates: Harness-Aware Routing

We've updated the Auto-Router to account for how Claude Code and Codex package requests. It now removes more harness context from classification, preserves encrypted delegated tasks, and shows exactly what the classifier received

Help shape the Auto-Router

Work with the LiteLLM team to test routing on your production traffic and help shape what we build next

Apply to Become a Design Partner

A coding harness sends more than the user's task. Requests also carry environment details, repository instructions, skill catalogs, and reminders. Those can dominate the classifier input or appear after the task as a separate message. Harness-aware routing uses the client identity and request format to find the task that needs a model

Update	What changes
Claude Code classification	Omits caller system text from the classifier input
Codex reminder handling	Removes recognized harness blocks while preserving the delegated task
Encrypted delegated tasks	Preserves encrypted task blocks in native Responses classifier calls
Classifier logs	Separates classifier input, masked source request, and classifier response

Less classifier input for Claude Code

In a Claude Code reproduction, a short binary-search question used 2,478 classifier input tokens. Environment details, agent definitions, and the skill catalog accounted for 7,613 characters in the classifier's user payload

For recognized Claude Code requests, the LLM classifier now omits caller system text. The selected model still receives the original system text, and the classifier keeps the current ask, configured prior turns, and conversation-depth signal

First classifier call	Before	After
User payload, characters	7,675	62
Total input tokens	2,478	524

That's about 79% fewer classifier input tokens on this call. The reproduction used Claude Code 2.1.268, a Haiku 4.5 classifier, and a three-turn session with one delegated subagent. It measures classifier overhead on that first call; total session cost and answer quality need separate evaluation. Reproduction and results

Task constraints supplied only in Claude Code system messages also stop influencing tier selection. Put constraints that should affect routing in the task itself

Keep the Codex task in view

Codex can append environment and repository instructions after a delegated task. Previously, that trailing message could become the classifier's current ask

For recognized Codex user agents, the router now strips complete harness blocks such as <environment_context> and <recommended_plugins>, plus the repository instruction envelope. The delegated task stays available for classification, and the selected model receives the original request

With classification_mode: user_turn and a session identifier, a fresh ask followed by reminder text gets classified. Subsequent assistant or tool continuations reuse its selected model without another classifier call. A new ask remains eligible for classification. Behavior and reproduction

Custom reminder_markers replace the built-in markers, so include every pair your harness needs when overriding them

Classify encrypted delegated tasks

Some Codex delegated tasks arrive in an agent_message containing encrypted_content. Converting that payload to ordinary text hid the task from the classifier and produced a SIMPLE verdict for a difficult request

The router now preserves the encrypted task in a native Responses classifier call. In the reproduction, the difficult task changed from SIMPLE on gpt-5.6-luna to REASONING on gpt-6-astra. The easy task stayed SIMPLE on gpt-5.6-luna, with both streaming and non-streaming requests tested. Reproduction and results

This requires a native OpenAI or Azure OpenAI Responses classifier deployment with credentials that can consume the encrypted content. Encrypted tasks bypass the local scoring shortcut in heuristic_first and hybrid modes. Unsupported deployments and decryption errors follow classifier_fallback. The live reproduction used OpenAI; Azure transport was not exercised

See what the classifier received

Open a request in Logs, then select its Classify row. New captures show Classifier input, Originating request, credentials masked, and Classifier response separately

This makes it possible to check whether the task reached the classifier and compare its verdict with the original request. Credentials, including Cookie and Set-Cookie headers, are masked in the source copy. Message-logging redaction settings still apply, and older rows retain their existing view. Logs update

Try it with your harness

Use a build containing the linked changes and send Claude Code or Codex traffic through the Auto-Router. Preserve the client's User-Agent header so the router can recognize it. Test Claude Code behavior through the real client: the browser routing preview has no client-identity field and uses generic classification behavior

Share what you find in the Auto-Router discussion