Compaction occurs way before context limit reached on GPT 5.4 #16333

Activity

1. GrahamCampbell

   added

   bugSomething isn't workingSomething isn't working

   on Mar 6, 2026
2. github-actions

   assigned

   rekram1-node

   on Mar 6, 2026
3. github-actions

   added

   coreAnything pertaining to core functionality of the application (opencode server stuff)Anything pertaining to core functionality of the application (opencode server stuff)

   on Mar 6, 2026
4. github-actions commented on Mar 6, 2026

   

   github-actionsBot

   on Mar 6, 2026 – with GitHub Actions

   Contributor

   More 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.
5. christallire commented on Mar 6, 2026

   

   christallire

   on Mar 6, 2026

   More actions

   @msadiks gpt 5.4 needs little more love here <3 thanks!
6. msadiks commented on Mar 6, 2026

   

   msadiks

   on Mar 6, 2026

   Contributor

   More actions

   @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.
7. oddko commented on Mar 6, 2026

   

   oddko

   on Mar 6, 2026

   Last edited by oddko

   More actions

   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.

   👍React with 👍7Reacted by Xin-Ma, RVZO6, Brendan Duke, herobrauni, Leon Liechti, Xynonners and Zhao Chengyu
8. GrahamCampbell commented on Mar 6, 2026

   

   GrahamCampbell

   on Mar 6, 2026

   ContributorAuthor

   More actions

   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.
9. oddko commented on Mar 6, 2026

   

   oddko

   on Mar 6, 2026

   More actions

   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.
10. msadiks commented on Mar 6, 2026

    

    msadiks

    on Mar 6, 2026

    Contributor

    More actions

    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.
11. vbuccigrossi

    added a commit that references this issue on Mar 7, 2026

    fix: resolve 8 open bug reports from upstream tracker

    ...

    0d1c221
12. JSCOP commented on Mar 10, 2026

    

    JSCOP

    on Mar 10, 2026

    Last edited by JSCOP

    More actions

    This is codex configuration setup

    model_reasoning_effort = "xhigh"
    model_context_window = 1050000
    model_auto_compact_token_limit = 900000
    approval_policy = "never"
    sandbox_mode = "danger-full-access"
    personality = "pragmatic"
    service_tier = "fast"
    

    This configuration setup makes context window to 1.05M (This is experimental feature but It works)
    I think opencode need to setup this configuration internally or change model loading features to load from outside configuration file for setup all other models.
13. rekram1-node commented on Mar 10, 2026

    

    rekram1-node

    on Mar 10, 2026

    Collaborator

    More actions

    This should be fixed if u run opencode models --refresh

    👍React with 👍3Reacted by Jonathan Vannier, JSCOP and Brendan Duke
14. oddko commented on Mar 11, 2026

    

    oddko

    on Mar 11, 2026

    More actions

    Looks like it's working, thanks a lot !
    
15. GrahamCampbell

    closed this as completedon Mar 11, 2026
16. JSCOP commented on Mar 11, 2026

    

    JSCOP

    on Mar 11, 2026

    More actions

    This should be fixed if u run opencode models --refresh

    This works! thanks :)

    👍React with 👍1Reacted by Zhao Chengyu