Composer 2 stopped halfway during code analysis and edits. It only provided the complete output after I asked it to continue working. I’ve seen this problem twice in a chat.
Steps to Reproduce
Prompt Composer 2 to make changes based on my requirements.
It stopped midway during thinking or code generation.
Expected Behavior
It should finish thinking and generating code in one go.
Hey, thanks for the report. From your screenshot, it looks like Composer 2 stopped mid-generation and only continued after you prompted it.
This can happen when the model hits its output token limit on complex tasks. It’s a known pattern where the model pauses and needs a nudge to keep going. The team knows about it and is tracking it.
To help us dig deeper, could you:
Next time this happens, grab the Request ID. Click the context menu (top right of the chat) and select Copy Request ID
Share it here so we can trace the exact session
Also, if you’re working with large files or complex prompts, try breaking your request into smaller, more focused tasks. This can help the model finish in one pass.
Thanks. From your reply, I probably know what had happened. I was asking Composer 2 to diagnose the storage size issue of a local database linked to an App I’m developing. A table inside the database is unusually large (4.5GB). I think this is probably the reason why the model reached its token limit very fast and stopped halfway.
Maybe Composer 2 is not suitable for this kind of task. I wonder if a flaship model like GPT 5.4 with a 1 million context window is able to handle it?
Table size (4,5 GB) doesn’t go straight into the model context. The model works with metadata, the schema, and requests, not raw data. So a 1M context window in GPT-5.4 won’t really change much here.
The early stop in Composer 2 is most likely about the output token limit, meaning how much the model can generate in one go, not how much it can read. This shows up more often on big multi-step tasks.
What will actually help:
Split the task into steps. First, analyze the table schema and estimate size by columns. Then, suggest optimization options, and so on.
For heavy diagnostics and planning, it can make sense to try Claude 4.6 Sonnet or GPT-5.4 in Max Mode. They handle long chains of reasoning better. Just keep in mind they come from the API pool and are billed by tokens.
If you really need to paste a big chunk of context, like schema plus sample rows plus logs, then yes, GPT-5.4’s large context window can help.
Composer 2 is fine for this kind of task, just break your prompts into smaller parts.
