Bad Request on Compaction only

[vmagalhaes]

Description

Hello,

When the compaction starts, it automatically stops saying "Bad Request" but only for Claude Opus 4.5 (I'm using github copilot integration).
And now that we are here, any idea why I can't use Opus 4.6 with my github copilot integration? It's enabled for my github enterprise integration.

Any idea why?

Plugins

oh-my-opencode

OpenCode version

1.1.65

Steps to reproduce

1. Start opencode
2. Write a prompt
3. Automatically starts the compaction
4. Bad Request

Screenshot and/or share link

Operating System

macOS 26

Terminal

Warp

[georgeharker]

Still occurring? I was having compaction related issues yesterday that appear to have been resolved on github's end: #13444

[vmagalhaes]

Still occurring? I was having compaction related issues yesterday that appear to have been resolved on github's end: #13444

Yes, the issue still persist, at least until 2h ago.

[dotned]

I keep getting this after a compaction, maybe the same issue. No logs tell me why and it is getting quite frustrating as it leaves the session unrecoverable.

Update:

Spent hours on this, trying older versions, even pulled the code and tried to get more insight. Got nowhere. I'm not even over quota...

I'm starting to think it's on the GH side at this stage.

Update 2:

I think there may be two issues here. @danriedl is showing a workaround to a >100% context usage issue, where it appears compaction overflows the models context window. My situation is different as I'm around 80% context usage and hitting errors.

I have managed to step past my issue by setting the compact agent to a non Anthropic model, GPT-5.2, which would suggest there is something in the way Claude models are called. I'm using GitHub Copilot Enterprise, specificity the GHE.com version.

I need to verify but some AI assisted analysis seems to point to packages/opencode/src/provider/transform.ts and specifically the code used to set the headers etc for Claude models, with GH rejecting with a 400.

[danriedl]

I do have the exact same issue. Was deleting .local/share/opencode but did not help.
Im on the newest release 1.2.17
Only way around this is to export the session after the context exceeded and compaction failed and continue in a new session.
Problem exists since 3 weeks, as @vmagalhaes also described.

Is it something hardcoded when the session should be compacted and it might changed?

Update:

This is what Opus4.6 reported for the opencode repository itself:

The Actual Bug in opencode
There are two bugs in the codebase:

1. compaction.ts sends all messages without checking if they fit the compaction model's context window. It should either truncate older messages or use a smarter strategy (e.g., summarize in chunks).
2. isOverflow doesn't apply the reserved buffer when limit.input is undefined (packages/opencode/src/session/compaction.ts:44-46). The else branch should be context - maxOutputTokens - reserved to ensure there's enough headroom for the compaction call itself.

Temporary Fix:
In your opencode.json, configure the compaction agent to use a smaller/cheaper model with a large context window:

{
  "agent": {
    "compaction": {
      "model": "github-copilot-enterprise/claude-sonnet-4"
    }
  }
}

[BleTib]

I also have the same issue with GHE on v1.2.22

[zwarag]

i also run into this issue

Temporary Fix: In your opencode.json, configure the compaction agent to use a smaller/cheaper model with a large context window:

{
"agent": {
"compaction": {
"model": "github-copilot-enterprise/claude-sonnet-4"
}
}
}

I'm using with github-copilot-enterprise/gpt-5.2-codex for compaction because I had the same issue with sonnet.

[tlaukkanen]

Having the same issue with GHE on v1.2.24 and Claude Opus 4.6. Switching to GPT-5.4 works and it does the compaction after saying "continue". Probably could then switch back to Opus 4.6 after compaction.

[vmagalhaes]

I'm closing the issue, looks like a common issue, for GHE it's preferred to use Copilot CLI (mainly because opencode doesn't manage copilot requests, only tokens).

[danriedl]

@vmagalhaes, opening an issue only to close it because you've ditched for another tool?
That's not how open source rolls. It's collaborative efforts and not a drive-by.
We've already nailed down the problem (check my response above), so this should stay open for someone with the guts to fix it.

[BleTib]

@vmagalhaes, opening an issue only to close it because you've ditched for another tool? That's not how open source rolls. It's collaborative efforts and not a drive-by. We've already nailed down the problem (check my response above), so this should stay open for someone with the guts to fix it.

agree

the workaround from above seems to work with gpt.

in ~/.config/opencode/opencode.json

{
  "agent": {
    "compaction": {
      "model": "github-copilot-enterprise/gpt-5.4"
    }
  }
}

[georgeharker]

That helps but you still pay a premium request on resuming after compaction