chore: implement streaming for anthropic.rs firstparty provider

[michaelneale]

Similar to openai: #3413

this adds in streaming for anthropic provider

[jamadeo]

Nice thanks @michaelneale ! when I run it the streaming looks good but my token count stays at 0 -- maybe not pulling the usage out correctly?

[jamadeo]

Turns out I had a similar issue with OpenAI: #3441

Is there a similar flag for Anthropic?

[michaelneale]

oh interesting, yes will take a look

[michaelneale]

adding in some logging... trying to see where the token counts are coming in:

🔍 Anthropic streaming line: data: {"type":"message_start","message":{"id":"msg_01VSXceb8PEy1o1DgnrvAuSv","type":"message","role":"assistant","model":"claude-sonnet-4-20250514","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":4,"cache_creation_input_tokens":3972,"cache_read_input_tokens":0,"output_tokens":1,"service_tier":"standard"}}            }
🔍 Anthropic streaming event: type=message_start, data_keys=["message"]
🔍 Anthropic streaming line:
◐  Yielding yaml yearnings...                                                                                                                                                                     🔍 Anthropic streaming line: event: content_block_start
🔍 Anthropic streaming line: data: {"type":"content_block_start","index":0,"content_block":{"type":"text","text":""}          }
🔍 Anthropic streaming event: type=content_block_start, data_keys=["content_block", "index"]
🔍 Anthropic streaming line:
🔍 Anthropic streaming line: event: ping
🔍 Anthropic streaming line: data: {"type": "ping"}
🔍 Anthropic streaming event: type=ping, data_keys=[]
🔍 Anthropic streaming line:
🔍 Anthropic streaming line: event: content_block_delta
🔍 Anthropic streaming line: data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"Here"}          }
🔍 Anthropic streaming event: type=content_block_delta, data_keys=["delta", "index"]
Here🔍 Anthropic streaming line:
🔍 Anthropic streaming line: event: content_block_delta
🔍 Anthropic streaming line: data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"'s a short joke for you:"}  }
🔍 Anthropic streaming event: type=content_block_delta, data_keys=["delta", "index"]
's a short joke for you:🔍 Anthropic streaming line:
🔍 Anthropic streaming line: event: content_block_delta
🔍 Anthropic streaming line: data: {"type":"content_block_delta","index":0,"delta":{"type":"text_delta","text":"\n\nWhy don't scientists trust atoms?"}   }
🔍 Anthropic streaming event: type=content_block_delta, data_keys=["delta", "index"]

according to docs:

https://docs.anthropic.com/en/docs/build-with-claude/streaming

The token counts shown in the usage field of the message_delta event are cumulative.

but am not seeing it - hrm... will have to put this on the back burner for a bit as not clear for a while. need to read some more.

if anyone knows and picks this up - go for it but will revisit it.

[michaelneale]

ok that took more time than I had hoped - it seems you have to accumulate/calculate as you go.

[jamadeo]

Nice, this worked for me.