Description

Summary

packages/llm/src/protocols/anthropic-messages.ts refuses to lower any assistant history containing a provider-executed (providerExecuted: true) tool-result whose tool name is not one of Anthropic's three native server tools (web_search, code_execution, web_fetch). Instead of degrading the unrecognized result, it fails the whole request with:

"Anthropic Messages does not know how to round-trip server tool result for "

This happens before any network call, during message lowering. The practical effect: once a conversation's history contains a provider-executed tool result that isn't one of those three (for example a provider-executed read), that conversation can no longer be continued on an Anthropic-API provider (e.g. Anthropic direct, Bedrock, Vertex). Switching provider mid-conversation — or otherwise replaying that history to an Anthropic route — is permanently blocked for that session.

Root cause

packages/llm/src/protocols/anthropic-messages.ts

The name mapper only knows three server tools (lines ~293-298):

const serverToolResultType = (name: string): AnthropicServerToolResultType | undefined => {
  if (name === "web_search") return "web_search_tool_result"
  if (name === "code_execution") return "code_execution_tool_result"
  if (name === "web_fetch") return "web_fetch_tool_result"
  return undefined
}

lowerServerToolResult throws on anything else (lines ~300-305):

const lowerServerToolResult = Effect.fn("AnthropicMessages.lowerServerToolResult")(function* (part: ToolResultPart) {
  const wireType = serverToolResultType(part.name)
  if (!wireType)
    return yield* invalid(`Anthropic Messages does not know how to round-trip server tool result for ${part.name}`)
  return { type: wireType, tool_use_id: part.id, content: part.result.value } satisfies AnthropicServerToolResultBlock
})

And it is reached for every provider-executed tool-result during lowering (lines ~461-463):

if (part.type === "tool-result" && part.providerExecuted) {
  content.push(yield* lowerServerToolResult(part))
  continue
}

So any providerExecuted: true tool-result with a name outside those three hard-fails the request.

Precedent

This is the Anthropic Messages analog of #29159 ("fix(llm): replay OpenAI Responses hosted tool history as native input items"), which was the same defect on the OpenAI Responses adapter — assistant history containing provider-executed hosted tool results was rejected during lowering with "only support text, reasoning, and tool-call content". That was fixed by teaching lowerMessages to lower the provider-executed pairs instead of rejecting them. The Anthropic Messages transform has the equivalent gap and never received the equivalent treatment.

Related: #45261 also concerns provider-executed tool calls breaking the rest of an Anthropic session (via repairToolCall), i.e. a nearby class of "provider-executed content is mishandled on the Anthropic path".

Suggested direction (not prescriptive)

Mirror #29159: in the Anthropic Messages transform, handle an unrecognized provider-executed tool-result by degrading it rather than throwing — e.g. lowering it to a normal content block (text, or a best-effort representation of the structured result) so the history remains replayable.
One caveat worth the maintainers' judgement: a native Anthropic tool_result block normally pairs with a client tool_use, which a provider-executed call does not have, so the exact safe representation (plain text vs. a synthesized block) needs deciding. The key ask is simply that an unrecognized provider-executed result should not hard-fail the entire request — a conversation should remain portable to an Anthropic provider.

Plugins

No response

OpenCode version

v2.0.18.

Steps to reproduce

1. Have a conversation whose assistant history contains a tool-result with providerExecuted: true and a tool name that is not web_search / code_execution / web_fetch (any provider/runtime that surfaces a locally/externally executed tool as provider-executed will produce one — e.g. a provider-executed read).
2. Continue that conversation on an Anthropic-API route (Anthropic, Bedrock, or Vertex Anthropic).
3. The request fails during lowering with Anthropic Messages does not know how to round-trip server tool result for — no network call is made.

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response