Anthropic Interaction Example

Back to Anthropic Messages.

This page mirrors the detailed Chinese interaction example in a compact English form. It is a protocol illustration, not a replayable upstream fixture.

Scenario

The user asks whether Beijing is suitable for running today. The client exposes a local get_weather tool, while the upstream may use an Anthropic server-side web_search tool.

1
client request
Client sends `messages`, local `get_weather`, server `web_search`, and `stream: true`.
2
provider pass-through
ProxAI forwards the Anthropic Messages request to the configured upstream.
3
first SSE turn
Upstream emits server-tool use/result blocks and a local `tool_use(get_weather)` block.
4
local tool runtime
Client runs `get_weather` using the accumulated `input_json_delta` fragments.
5
second request
Client sends the previous assistant `tool_use` plus a user-side `tool_result`.
6
final SSE answer
Upstream emits text deltas and ends with `message_delta(stop_reason=end_turn)` plus `message_stop`.

ID relationships

ID	Source	How it is used later
`srvu_01`	`server_tool_use`	Executed by the upstream service; the client does not send a matching tool result.
`srvr_01`	`web_search_tool_result`	Identifies the server-tool result block itself.
`toolu_01`	`tool_use(get_weather)`	Client executes the local tool and sends `tool_result.tool_use_id = toolu_01` in the next request.

Stream assembly rule

Anthropic streaming joins content deltas by content_block index:

content_blocks[index] = started block
partial_tool_inputs[index] += input_json_delta.partial_json

When content_block_stop(index=2) arrives, the client can parse the accumulated partial JSON and use the same block’s ToolUseBlock.id as the local tool call id.