Code generation frequently stops mid-way due to output token limits

[enteve]

Description:

When using DeepAgent to generate code, I frequently encounter an issue where the generation process stops abruptly due to the LLM's output token limit. This often happens before the code is complete, leaving partial or unusable results.

I've tried adding prompts instructing the agent to generate code in segments/batches, but the issue persists.

Has anyone experienced this or found a workaround? Some potential solutions I'm considering:

1. Automatic resumption when output is truncated - detect incomplete code and continue generation
2. Built-in chunking strategy for large code generation tasks
3. Configurable token limit awareness in the agent

Questions:

• Is there a recommended way to handle long code generation?
• Are there any built-in features I might be missing to mitigate this?
• Would it be possible to add support for streaming/chunked code generation?

Any guidance would be appreciated!

[christian-bromann]

enteve thanks for reporting, I experienced the same and will take a look.

[christian-bromann]

enteve which deepagents version are you using?

[christian-bromann]

Also, mind sharing the exact error message? I am getting this:

{"type":"error","error":{"type":"invalid_request_error","message":"prompt is too long: 219536 tokens > 200000 maximum"},"request_id":"req_011CY2kPwWtNt2aoncRPXh3B"}

[enteve]

"deepagents": "^1.7.3"

i didn't keep the error message, i'll share the next time. it's not the prompt that exceeds the length. it's the output token. the error message contains something like

{"finish_reason": "length"}

[arielb-oasis]

Is there any solution? I dropped deepagents and went to Claude Claude because of limits I got each time even with file system middleware

[christian-bromann]

Yes arielb-oasis we have a solution in #234 which we will ship tomorrow