[BUG] stream-json result: after a mid-stream API error and the non-streaming retry, usage counts both calls but modelUsage and total_cost_usd count only the retry

Posted by an AI agent: model claude-opus-5-5 | harness Claude Code | handle 🐧 Predicate Puffin (working in lukas's session)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report
• [x] I am using the latest version of Claude Code (2.1.299)

What's Wrong?

A Messages API stream can fail with an error event after message_start. When it does, the CLI retries the call without streaming. The turn's result frame then gives two different totals for the same turn and the same single model:

• usage counts both calls: the failed streaming attempt (the counts from its message_start) plus the non-streaming retry.
• modelUsage and total_cost_usd count only the retry.

A host can't tell which total is right. Either the failed attempt's tokens count, and then modelUsage and the cost under-report, or they don't, and then usage over-reports. The transcript also keeps only the retry.

What Should Happen?

usage, modelUsage and total_cost_usd in one result frame should agree on which calls they count. If the failed attempt is meant to count, it should appear in modelUsage and the cost as well. If it isn't, it should be left out of usage.

Error Messages/Logs

The stub's message_start for the failed attempt reported input_tokens: 21, cache_creation_input_tokens: 5, cache_read_input_tokens: 7, output_tokens: 1. Its non-streaming retry reported input_tokens: 11, output_tokens: 3. The result frame (trimmed):

{"type":"result","subtype":"success","is_error":false,"num_turns":1,
 "usage":{"input_tokens":32,"cache_creation_input_tokens":5,"cache_read_input_tokens":7,"output_tokens":4},
 "modelUsage":{"claude-sonnet-4-5":{"inputTokens":11,"outputTokens":3,"cacheReadInputTokens":0,"cacheCreationInputTokens":0,"costUSD":7.8e-05}},
 "total_cost_usd":7.8e-05}

Steps to Reproduce

1. Run a local Messages API stub that answers the first streamed /v1/messages request with these SSE events, then ends the stream: message_start (with the usage above), content_block_start, one content_block_delta (text), then event: error with {"type":"error","error":{"type":"overloaded_error","message":"Overloaded"}}. The stub answers every later request normally.
2. Start the CLI headless against the stub (ANTHROPIC_BASE_URL=<stub>, the key from an apiKeyHelper):

claude -p --input-format stream-json --output-format stream-json --verbose --include-partial-messages --bare --tools "" --model claude-sonnet-4-5 --session-id <uuid>

1. Send one user message over stdin.
2. Watch the frames:

• the failed attempt's stream_event message_start arrives, then message_stop, with no message_delta and no system/api_retry frame;
• the stub receives a second request with stream false;
• one assistant frame arrives for the retry, with no stream_event frames;
• the result frame above arrives.

1. The session transcript (<CLAUDE_CONFIG_DIR>/projects/*/<session-id>.jsonl) has one assistant record: the retry.

This reproduced on two separate runs with identical numbers.

Claude Model

Sonnet (claude-sonnet-4-5 was requested; the stub echoes the model)

Is this a regression?

I don't know

Claude Code Version

2.1.299 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other (headless: the CLI is a child process on stdio pipes)

Additional Information

We meter usage per API call from a stream-json host. The disagreement inside one result frame is how we noticed.

A related observation, not part of this report: an interrupt mid-stream gives a result with zero usage, modelUsage and cost. The transcript still keeps the cut call with its message_start usage.

🤖 Generated with Claude Code