anthropic stream ended before message_stop error

Not specified by the original author.

TL;DR

Install the SDK: npm install @anthropic-ai/sdk. In the route handler, set the SSE response headers BEFORE creating the stream: Content-Type of text/event-stream, Cache-Control of no-cache, and Connection of keep-alive.

Steps

1. Install the SDK: npm install @anthropic-ai/sdk. The client-SDK docs list TypeScript support for Node.js, Deno, Bun, and browser.
2. In the route handler, set the SSE response headers BEFORE creating the stream: Content-Type of text/event-stream, Cache-Control of no-cache, and Connection of keep-alive. Headers sent after the first write throw.
3. Start the message stream and iterate events. Forward only text_delta payloads as event lines. Skip message_start, content_block_start, content_block_stop, and message_delta; they carry no user-visible text.
4. Handle unknown event types gracefully with a skip branch. Anthropic may add new event types under its versioning policy; your endpoint must not crash when one appears.
5. On message_stop, end the response exactly once. On a client disconnect, abort the upstream stream so you stop paying for tokens nobody reads.
6. Wrap the whole handler in error handling that ends the response on failure. A mid-stream exception with no end call leaves the browser hanging until timeout.
7. Keep the API key in server config. The TypeScript SDK runs in the browser too, which makes it tempting to call from frontend code; do not. The endpoint is the trust boundary.
8. Failure modes this prevents: setting headers after the first write and crashing; forwarding non-text events as visible output; orphaned streams burning tokens after the user navigated away.

When to use

You are seeing this: The client-SDK docs list TypeScript support for Node.js, Deno, Bun, and browser. Use this skill when you run into "Anthropic streaming on Express: set SSE headers first, forward textdelta, close on messagestop".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Why this happens

The original report does not dig into a root cause. It documents the symptom and the fix that resolved it.

The client-SDK docs list TypeScript support across Node.js, Deno, Bun, and browser, and the streaming reference documents the event flow and the graceful-unknown-event-type rule. Verified live at https://docs.anthropic.com/en/api/client-sdks and https://docs.anthropic.com/en/api/streaming.

Not specified by the original author.

TL;DR

Do not concatenate deltas across different indices into one string; each index is a separate content block (text vs tool_use vs thinking). Anthropic's versioning policy says new event types may be added, so a skip branch for unrecognized events is required, not optional. FastAPI: wrap the async event iterator in a StreamingResponse with the media type set to text/event-stream, and yield server-sent-event formatted lines as you consume deltas.

When to use

You are seeing this: Start the stream with the SDK's streaming interface, e.g. client.messages.stream(...) as a context manager. Use this skill when you run into "Anthropic streaming on Python: accumulate text_delta per block, serve via FastAPI StreamingResponse or a Flask generator".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Streaming Claude output from Python (FastAPI and Flask)

1. Start the stream with the SDK's streaming interface, e.g. client.messages.stream(...) as a context manager. The docs name the helpers that accumulate the full message for you: stream.get_final_message() in Python and .finalMessage() in TypeScript.
2. Handle the documented event flow per content block: message_start carries a Message object with empty content; then per block a content_block_start, one or more content_block_delta events, and a content_block_stop; each block's index matches its index in the final Message content array; the stream ends with message_delta then message_stop.
3. Accumulate text_delta text per block index. Do not concatenate deltas across different indices into one string; each index is a separate content block (text vs tool_use vs thinking).
4. Ignore unknown event types gracefully. Anthropic's versioning policy says new event types may be added, so a skip branch for unrecognized events is required, not optional.
5. FastAPI: wrap the async event iterator in a StreamingResponse with the media type set to text/event-stream, and yield server-sent-event formatted lines as you consume deltas.
6. Flask: return a Response wrapping a generator that yields the same event lines. Flask's sync view pairs with the SDK's sync streaming interface.
7. Flush headers before the first token arrives so the browser's event source connects while the model is still generating. On any mid-stream exception, end the stream cleanly instead of leaving the connection hanging.

Failure modes this prevents: mixing deltas from different content-block indices into garbled text; crashing on a new event type after an API update; buffering the whole response and defeating the purpose of streaming.

The streaming reference documents the messagestart / contentblockstart / contentblockdelta / contentblockstop / messagedelta / messagestop flow, per-block indices, the getfinal_message and finalMessage helpers, and the rule to handle unknown event types gracefully. Verified live at https://docs.anthropic.com/en/api/streaming and https://docs.anthropic.com/en/docs/build-with-claude/streaming.

Not specified by the original author.

TL;DR

The message stop\reason is max\tokens instead of end\turn. Do not confuse this with a rate limit: OTPM counts actual generated tokens, and maxtokens itself does not count against OTPM.

Steps

1. Responses end mid-thought. The message stop_reason is max_tokens instead of end_turn.
2. stop_reason: max_tokens means the turn hit the output ceiling. With thinking on, check whether the thinking budget consumed most of it: thinking tokens count toward max_tokens for the turn, so a 16k budget against a 20k maxtokens leaves almost nothing for the answer. The budget must be less than `maxtokens` (interleaved thinking is the one exception, where the budget can exceed it).
3. Raise max_tokens to leave headroom above the thinking budget plus the expected answer length.
4. Or lower the thinking budget; the budget is a target, not a guarantee of quality.
5. Do not confuse this with a rate limit: OTPM counts actual generated tokens, and max_tokens itself does not count against OTPM.

When to use

You are seeing this: Responses end mid-thought. Use this skill when you run into "Anthropic answer cut off mid-sentence: stopreason maxtokens, and thinking budget eats the same budget".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Why this happens

stop\reason: max\tokens means the turn hit the output ceiling.

stopreason values, the budgettokens must be less than maxtokens rule, and thinking counting toward maxtokens are documented on the streaming and extended thinking pages. Verified live at https://docs.anthropic.com/en/api/messages-streaming and https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking.