[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

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.

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

@thdxr

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

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

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

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

@thdxr

@thdxr I hit it every 30 minutes )

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

https://opencode.ai/s/WjniS1OH

@thdxr

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

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

this is rendering opencode unusable for me for even the most basic tasks, Context: building a new app package within a large monorepo.

@oscaromsn

@oscaromsn I understand you pain! But even with this bug code made with opencode have ijigen-level quality

@flybayer

@flybayer we need help @thdxr to fix it, all other tool in code agent space no have so high code quality. I try few in parallel with same codebase/prompts/model

every time i hit this, i have to kill and restart opencode. is there a better way? it gets this error when it tries to read a large webpage or largish (600k) file . also, would be nice if i could cut and paste the error instead of a screenshot, but copy /paste doesn't seem to work in opencode

For those looking for a workaround right now. I'm making very fine grained checklists to work through and put specific instructions in AGENTS.MD to stop and ask if you are ready to continue after checking off each item. It lets me manually eyeball the current context number and if its about to blow I get it to update where its up to in the checklist and restart a new session to continue.

I know this sounds cumbersome but its better than losing the whole chat.

The other challenge that happens when one of these errors pops is that you lose the last prompt you wrote so if it was very long it means a full rewrite of that one.

another strategy is to make the agent use a markdown file to store its plan and progress. that way you can continue in new sessions or if a session fails. also use compact at each checkpoint.

we upgraded to ai sdk v5 which paves the way to fix this issue - note there is no magic fix, what i'm planning to do here is run compaction in between if the context gets too big in the middle of a loop

looks good

every time i hit this, i have to kill and restart opencode.

When this happens to me I run /compact manually and I can continue the session.

To stay organized issues are automatically closed after 90 days of no activity. If the issue is still relevant please open a new one.

feat: agent compression, inspection logic, governance hooks (anomalyc…

feat: agent description compression with lazy loading (anomalyco#696)

fix: allow data:image/* URLs in markdown renderer (anomalyco#491)

Navigation Menu

Repository navigation

[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 #491

4 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[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 #491

Description

Activity

ezynda3 commented on Jun 28, 2025

thdxr commented on Jun 29, 2025

joegearlesss commented on Jun 29, 2025

thdxr commented on Jun 29, 2025

thdxr commented on Jun 29, 2025

thdxr commented on Jun 29, 2025

joegearlesss commented on Jun 29, 2025

joegearlesss commented on Jun 29, 2025

joegearlesss commented on Jun 29, 2025

oscaromsn commented on Jul 2, 2025

4 remaining items

flybayer commented on Jul 4, 2025

joegearlesss commented on Jul 5, 2025

joegearlesss commented on Jul 5, 2025

stellarthemes commented on Jul 7, 2025

hbthegreat commented on Jul 9, 2025

sockthedev commented on Jul 9, 2025

thdxr commented on Jul 9, 2025

joegearlesss commented on Jul 9, 2025

BernieSumption commented on Jul 10, 2025

thdxr commented on Mar 25, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions