Compaction occurs way before context limit reached on GPT 5.4

[GrahamCampbell]

Description

GPT 5.4 supports a ~1M context size, however compaction is occurring way before this, similar to the levels at 5.2.

Plugins

No response

OpenCode version

1.2.20

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

macOS 26.3.1

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• How to enable 1M context in OpenCode for GPT-5.4? Compaction triggers at 272k, not 1M #16308: Same issue reported same day — GPT-5.4 compaction triggers at 272k instead of the model's 1M context window
• Auto-compaction triggers at ~200k tokens instead of model's actual context window (1M) #15871: Auto-compaction triggers at ~200k tokens instead of model's actual context window (1M)
• GPT 5.2 Codex - Compaction happens way too early and context limit is wrong #11086: GPT 5.2 Codex — compaction happens way too early and context limit is wrong (predecessor for GPT 5.2)
• GPT models consistently run out of context earlier than indicated #12621: GPT models consistently run out of context earlier than indicated
• Model context limits not respected #11286: Model context limits not respected

If your issue is already covered by one of the above, please consider closing this one and adding a "+1" or comment there instead.

[christallire]

@msadiks gpt 5.4 needs little more love here <3 thanks!

[msadiks]

@christallire @GrahamCampbell
Adding the model to allowance list is not related to compaction and the model context window config in models.dev was also correct the last time I checked it. The issue is probably the compaction logic doesn't respect the context window or might be a calculation error. I may take another look at this when I am home.

[oddko]

I think the feature (extended context window) is also opt-in/beta-ish in the first-party Codex app and burns 2X usage beyond the 272K limit. So by default you'll get hit by compaction when reaching 272K, regardless of which software you use.

If you can implement this I'd also be interested in the fast mode ("/fast" in Codex CLI/App) toggle/config if you can.

[GrahamCampbell]

I wonder if the usage-API bypasses that limit. Their documentation there says 1.05M and doesn't say any special headers or config is needed... though that could just be crappy docs.

[oddko]

On Codex (apps) it requires modifying ~/.codex/config.toml so perhaps yeah somewhat undocumented stuff, but just like the fast mode it's probably sent through the headers.

[msadiks]

As I reviewed the compaction logic, it basically limits the compaction window to model input limit minus reserved limit to be able to get a summary of the history. The problem with Openai oAuth provider is that it has an input window of 272k and a context window of 1m.

So to actually use 1m context window the options I can think of are:

1. disable the compaction entirely.
2. use gpt-5.4 from another provider (like from opencode zen with 900k input window)
3. refactor the compaction logic to use another strategy.
4. configure the current compaction logic to use compacting agent input limits instead of current agent limits and set a compact agent with a larger input limit. considering the input limit of current agent since the summary goes there.

In my opinion, first option wouldn't be beneficial at all. Third and fourth options are long term and need a deeper analysis. Best short term solution here is either going with second option or using the model as it is.

I am not a maintainer of the repo so this answer is based on a quick review of the codebase.