Issues using compactionControl

[gustavosizilio]

Hello, im using the beta.messages.toolRunner with streaming mode. Im sure that the streaming is enabled (true). Everything was working fine until I enabled the compactionControl

I just added the param:

compactionControl: {
enabled: true,
contextTokenThreshold: 10000,
}

When I enable the compactionControl I always get the following error when the lib is trying to use (inside the lib):

const isCompacted = await this.#checkAndCompact();

ERROR:

AnthropicError: Streaming is required for operations that may take longer than 10 minutes. See https://github.com/anthropics/anthropic-sdk-typescript#long-requests for more details
at Anthropic.calculateNonstreamingTimeout (/Users/gustavosizilio/Workspace/x/node_modules/@anthropic-ai/sdk/src/client.ts:834:13)
at Messages.create (/Users/gustavosizilio/Workspace/x/node_modules/@anthropic-ai/sdk/src/resources/beta/messages/messages.ts:110:30)
at BetaToolRunner._BetaToolRunner_checkAndCompact (/Users/gustavosizilio/Workspace/x/node_modules/@anthropic-ai/sdk/src/lib/tools/BetaToolRunner.ts:124:54)
at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
at BetaToolRunner.[Symbol.asyncIterator] (/Users/gustavosizilio/Workspace/x/node_modules/@anthropic-ai/sdk/src/lib/tools/BetaToolRunner.ts:201:31)
at ParsePromptStream.parse (/Users/gustavosizilio/Workspace/x/src/app/ai/chat-agents/parsers/parse-prompt-stream.ts:40:24)
at ChatAgentsService.wssChatAgentsNewMessageHandler (/Users/gustavosizilio/Workspace/x/src/app/ai/chat-agents/chat-agents.service.ts:360:13)
at EventSubscribersLoader.wrapFunctionInTryCatchBlocks (/Users/gustavosizilio/Workspace/x/node_modules/@nestjs/event-emitter/lib/event-subscribers.loader.ts:190:14)

Considerations:

One detail is that this is giving this message even though the conversation is less than a minute... im trying with compaction threshold very small to test the functionality.

[gustavosizilio]

This is not supposed to use streaming from the original conversation configuration? (Maybe it should reuse all the configuration?)

[gustavosizilio]

What I "believe":

• The SDK calculates expected time using this formula:
  expectedTime = (60 minutes × maxTokens) / 128000

• With your 64000 tokens:
  expectedTime = (60 × 64000) / 128000 = 30 minutes

• Since 30 minutes > 10 minutes, the SDK throws the error immediately before even making the request. It doesn't matter how fast the actual call would be.

[gustavosizilio]

When I reduce the tokens I dont get the timeout error... but I get another issue as you can see below. This issue seems to be similar to many others where the SDK exposes extra params to the messages (many issues reporting this on the github issue tracker)

Disclaimer: Again, im just guessing here and trying to help.

[gustavosizilio]

My conclusion is that if your context is using beta header, you need to compact using the same headers otherwise you will have issues.

Also, you need to use stream if your mac tokens is big.

I implemented my own compaction method because of these two issues.

[Zelys-DFKH]

Traced the issue. The #checkAndCompact() method calls this.client.beta.messages.create() without streaming, passing the full max_tokens from the tool runner params. The SDK's calculateNonstreamingTimeout throws when max_tokens exceeds ~21K (or 8192 for claude-opus-4 models), since it estimates the request could take longer than 10 minutes.

The fix is to cap max_tokens at a small constant (e.g. 4096) for the compaction summary request — a summary doesn't need the full output budget. Happy to open a PR if that sounds right.