Translated route drops tool_reference blocks from tool results, so ToolSearch returns empty and routed models call it in a loop #6647

Client or integration

Claude Code

Area

Tools / MCP / web search

Summary

Claude Code's ToolSearch tool answers with a tool_result whose content is only tool_reference blocks ([{"type":"tool_reference","tool_name":"Bash"}, …]). On the translated Claude Messages → Responses route, toolResultOutput in src/claude/inbound.ts keeps text, image and document items and drops everything else, so the routed model receives an empty function_call_output for every ToolSearch call. It concludes the tools did not load and calls ToolSearch again; one subagent made 26 ToolSearch calls in 9 minutes, restarting its plan after each.

Expected: the routed model sees which tools were loaded, as it does for any other tool result. src/claude/compatibility.ts already lists tool_reference as a supported result block type, so the translation is expected to carry it.

When it happens. Claude Code only keeps tool search on when it believes it is talking to Anthropic. That is the case in the first-party desktop intercept mode (desktopMode: "first-party"), where ANTHROPIC_BASE_URL is not set. #4838 covers the opposite configuration (gateway mode turns tool search off). In first-party mode the deferral stays on and tool_reference results do flow through the translation, where they are dropped.

Fix that works here. In toolResultOutput, map a tool_reference item to text:

} else if (item.type === "tool_reference" && typeof item.tool_name === "string") {
  out.push({ type: "input_text", text: `Tool loaded: ${item.tool_name}` });
}

With that one branch the routed model loads its tools once and moves on.

Reproduction

1. Run opencodex with a Claude Code route translated to a Responses model (here a combo whose first member is meta-muse/muse-spark-1.3-contributor).
2. POST /v1/messages with a history containing an assistant tool_use for ToolSearch (input {"query":"select:WebFetch"}), followed by a user tool_result for it with content [{"type":"tool_reference","tool_name":"WebFetch"}], then a user text message: Repeat the tool result you received, verbatim.
3. Observe: the model reports an empty tool result. With the fix above, it answers Tool loaded: WebFetch.

Version

2.76.0 (also present in 2.78.0-preview.20261005)

Operating system

macOS 26.6.2 (arm64); also macOS 26.7.1 (arm64)

Provider and model

meta-muse / muse-spark-1.3-contributor (via combo); not provider-specific, the drop happens before routing

Logs or error output

No response

Screenshots and supporting files

No response

Redacted configuration

No response

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.