[v0.1.162] [Anthropic: Claude Sonnet 4]: AI_APICallError: input length and 'max_tokens' exceed context limit: 140740 + 64000 > 200000, decrease input length or 'max_tokens and try again

[joegearlesss]

https://opencode.ai/s/aBPMCaSP

[ezynda3]

Having the same issue. The default max output used to be 50000 for claude no? Context runs out way to quickly when set to 64k I think.

[thdxr]

so it turns out if you specify max tokens it needs to be less than (max context - input) but it's hard to know the exact input up front

thinking about how to deal with this

[joegearlesss]

@thdxr if you tell me some specific way to share debug information, I'll do it

[thdxr]

the problem might be that auto compaction isn't happening - checking now

[thdxr]

we have the following logic

      const tokens =
        previous.metadata.assistant.tokens.input +
        previous.metadata.assistant.tokens.cache.read +
        previous.metadata.assistant.tokens.cache.write +
        previous.metadata.assistant.tokens.output
      if (
        model.info.limit.context &&
        tokens >
          Math.max(
            (model.info.limit.context - (model.info.limit.output ?? 0)) * 0.9,
            0,
          )
      ) {
        await summarize({
          sessionID: input.sessionID,
          providerID: input.providerID,
          modelID: input.modelID,
        })
        return chat(input)
      }

for anthropic it should be (200,000 - 64,000) * 0.9 = 122,400

so i'm not sure how it got to 140,740

[thdxr]

oh i see the issue it only does the compaction check when you send a message, if it's running in a loop for a while it can exceed it and not trigger compaction

to fix this we need to upgrade to v5 ai sdk

[joegearlesss]

@thdxr I hit it every 30 minutes )

[joegearlesss]

and /compact do not work for this case. restart and reopen session do not working for this case too

https://opencode.ai/s/WjniS1OH

[joegearlesss]

@thdxr if you can give me some way to continue sessions like this before redesign or update to v5, it will be great

[oscaromsn]

same situation here when asking the agent to fetch some web content. currently my main reason to not use opencode more.