GPT 5.2 Codex - Compaction happens way too early and context limit is wrong

[Simplereally]

Description

GPT 5.2 Codex via Copilot is compacting at 100k tokens when the model's context is 272k.

Plugins

None

OpenCode version

latest

Steps to reproduce

1. Use gpt-5.2-codex via github copilot
2. Context in top right indicates max 100k when it is actually 272k
3. Reach ~90k+ tokens in session and compaction happens

Screenshot and/or share link

Operating System

Windows 10 latest

Terminal

Windows Terminal

[github-actions]

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

• Default Codex context length not matching Pro plan #7705: Default Codex context length not matching Pro plan - Reports that GPT 5.2 Codex shows 400k but Pro plan actually has 250-260k, causing early compaction
• GitHub Copilot GPT-5 throwing "invalid_request_body" on requests after compaction #10650: GitHub Copilot GPT-5 throwing "invalid_request_body" on requests after compaction - Reports context compaction failures with GitHub Copilot GPT-5
• github copilot: 2x token (counts total=total+cached) compacts at 64k, instead of 128k #3314: github copilot: 2x token (counts total=total+cached) compacts at 64k, instead of 128k - Reports incorrect token counting and premature compaction with GitHub Copilot models
• Compaction overflow check doesn't account for large tool outputs until the next step, causing context overflow #10634: Compaction overflow check doesn't account for large tool outputs until the next step, causing context overflow - Reports compaction being triggered too early due to incomplete overflow checks

Feel free to ignore if none of these address your specific case.

[Deftera186]

This is probably related to a problem in models.dev that was fixed 2 weeks ago:
anomalyco/models.dev@b0e5b91

maybe you still have the old cache for some reason? try to remove your ~/.cache/opencode/models.json and update if it helps

[Waishnav]

Yup, they have fixed this issue. It is working all fine

[dlukt]

Nope. Deleting ~/.cache/opencode/models.json does not fix it.

49,590  39% ($0.00)
v1.1.51

50k tokens and already 39% context used.

grep github-copilot/gpt-5.2 ~/.cache/opencode/models.json

No results.

The search for github-copilot in this repo reveals no results either.

How can you greatly announce your partnership when you don't even do the basics?

[dlukt]

The issue is most likely because you source https://models.dev/api.json
which incorrectly states the limits for context and output to 128k and 64k

I'm letting OpenAI/gpt-5.2 now fix this issue and will send a PR

[dlukt]

"provider": {
    "github-copilot": {
      "models": {
        "gpt-5.2": {
          "limit": {
            "context": 400000,
            "input": 272000,
            "output": 128000
          }
        },
        "gpt-5.2-codex": {
          "limit": {
            "context": 400000,
            "input": 272000,
            "output": 128000
          }
        }
      }
    },
    "github-copilot-enterprise": {
      "models": {
        "gpt-5.2": {
          "limit": {
            "context": 400000,
            "input": 272000,
            "output": 128000
          }
        },
        "gpt-5.2-codex": {
          "limit": {
            "context": 400000,
            "input": 272000,
            "output": 128000
          }
        }
      }
    }
  }
}