[terence71-glitch]

Summary

This implements semantic passthrough for valid Claude Code /v1/messages payloads on direct Claude-Code-style routes:

• native Claude OAuth provider: claude
• Claude Code-compatible relay providers: anthropic-compatible-cc-*
  For verified Claude Code / Claude CLI requests, OmniRoute now avoids broad message normalization that can rewrite or reorder the client-provided Anthropic Messages payload before it reaches an upstream that already understands Claude Code semantics.
  Closes [Feature] Preserve Claude Code /v1/messages payload semantics on direct Claude-Code paths #2347

What changed

• Added narrow Claude Code semantic passthrough detection for direct Claude-Code paths.
• Detects Claude Code / Claude CLI requests using indicators such as:
  • claude-code
  • claude-cli
  • x-app: cli
  • x-claude-code-session-id
• Skips broad normalizeClaudeUpstreamMessages() for verified Claude Code requests on the native claude path.
• Preserves Claude-native messages[] through the anthropic-compatible-cc-* relay path before required relay-specific transforms.
• Keeps non-Claude-Code Claude-shaped requests on the existing normalization/repair path.
• Keeps model resolution, credential selection, native Claude OAuth behavior, compatible relay URL/path/header/signing behavior, and response handling intact.

Why

Claude Code already emits valid Anthropic Messages API payloads for coding tasks, including message ordering, tool-use/tool-result structure, system blocks, cache-control placement, and thinking/context fields.
Broad normalization on direct Claude-Code-style routes can unnecessarily change that structure by:

• moving role: "system" messages into top-level system
• removing system messages from messages[]
• filtering empty text blocks
• converting file/document-like blocks into text
• dropping unknown-but-valid content block types
• changing tool-use/tool-result history shape
  This PR keeps the client-provided messages[] semantics intact for valid native Claude Code traffic while still allowing OmniRoute to perform routing, model resolution, OAuth/account handling, headers/signing, and provider-specific compatibility work.

Scope

This is semantic passthrough, not raw byte-for-byte passthrough.
OmniRoute still parses JSON and may still apply required behavior such as:

• resolving the final upstream model id
• selecting credentials/accounts
• applying native Claude OAuth headers/signing behavior
• choosing the Claude Code-compatible relay path for anthropic-compatible-cc-*
• preserving required relay builder behavior
• performing targeted compatibility work where required
  This does not change general cross-provider translation behavior.

Acceptance criteria covered

• Valid native Claude Code /v1/messages requests to claude keep messages[] unchanged through chatCore translation.
• Valid native Claude Code /v1/messages requests to anthropic-compatible-cc-* keep messages[] unchanged before required relay transforms.
• Non-Claude-Code Claude requests still use existing normalization behavior.
• Native Claude Code requests still receive the resolved upstream model.
• Claude Code-compatible relay requests still use the compatible relay builder and /v1/messages?beta=true-style path behavior.
• Existing OpenAI/Gemini/non-CC translation paths are unaffected.

Tests

node --import tsx --test --test-force-exit tests/unit/chatcore-translation-paths.test.ts tests/unit/claude-code-compatible-request.test.ts tests/unit/chatcore-sanitization.test.ts
Result:
tests 87
pass 87
fail 0
Notes
npm run typecheck:core is currently blocked by pre-existing unrelated errors in open-sse/services/contextManager.ts.

[gemini-code-assist]

Code Review

This pull request introduces a 'semantic passthrough' mechanism for Claude Code and CLI requests, allowing them to bypass standard message normalization and preserve complex content blocks like documents and tool results. It includes detection logic based on headers and User-Agent strings, updates to the translation service to handle preserved messages, and comprehensive unit tests. A review comment identified an opportunity to optimize performance by removing redundant deep cloning of the messages array during the semantic body preparation.

[gemini-code-assist]

The messages array is being deep-cloned twice when preserveClaudeMessages is enabled. First, prepareClaudeCodeCompatibleSemanticBody performs a deep clone of the entire array. Subsequently, cloneClaudeCodeCompatibleMessagesFromClaude (called from buildClaudeCodeCompatibleRequest) iterates over this already-cloned array and performs another deep clone on each individual message object. For large conversation histories typical in Claude Code sessions, this redundant deep cloning can impact performance and memory usage.

Suggested change

	messages: Array.isArray(claudeBody.messages)
	? (cloneValue(claudeBody.messages) as Array<Record<string, unknown>>)
	: [],
	messages: Array.isArray(claudeBody.messages)
	? (claudeBody.messages as Array<Record<string, unknown>>)
	: [],

[diegosouzapw]

Thank you @terence71-glitch for this excellent contribution! 🎉

Your semantic passthrough implementation is exactly what Claude Code users need — preserving the carefully structured messages[] array (document blocks, tool_use/tool_result chains, cache_control placement) for native Claude-Code routes while keeping the existing normalization path intact for everything else.

The 87-test coverage is outstanding, and the narrowly scoped detection logic (User-Agent + x-app + session-id headers) is clean and robust.

This has been merged into release/v3.8.0 and will ship in the upcoming release. Thanks again for the quality work! 🚀