Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

The following comment was made by an LLM, it may be inaccurate:

I found one potentially related PR:

PR #9656: feat(opencode): trigger compaction earlier and add multi-file read
#9656

This PR appears to address compaction triggering behavior, which is related to the compaction threshold configuration being added in the current PR (#10123). However, it seems to focus on multi-file read functionality as well, so they may not be exact duplicates. You should review #9656 to ensure there's no overlap in the compaction logic implementation.

EDITS: Correctly mentioned the issue being fixed.

Reviewed the potentially related PR, this one focuses solely on customizing the early compaction threshold as per the user's need, while preventing infinite compaction loop. Meanwhile the related PR (#96956) is a hardcoded early compaction (at 90% of the context) fix along with multi-file read validation.

@jeremyakers Just +1 on this PR! Custom compaction thresholds are critical for long-context models like Gemini/GPT that degrade quality at 30-50% context. The hardcoded 75% threshold causes severe performance issues in production workflows.

Having per-model threshold configuration would allow power users to:

• Set Gemini to compact at 30% (300k tokens) instead of 75%
• Set Claude to compact at 50% instead of 75%
• Keep shorter-context models at default 75%

This aligns perfectly with the 'auto-archive' mode you mentioned - both give users fine-grained control over when context gets compacted.

Happy to test beta implementations!

@WietRob I have added a commit allowing model-specific threshold control, thank you for the suggestion!

Here is a sample config.

{
  "$schema": "https://opencode.ai/config.json",
    "compaction": {
        "token_threshold": 100000,
        "context_threshold": 0.30,
        "models": {
            "opencode/minimax-m2.1-free": {
                "token_threshold": 50000,
                "context_threshold": 0.10
            }
        }
    }
}

It would be great help if you could test it out as well.

@saksham1341 Tested this with opencode/minimax-m2.1-free using a 5% threshold in my config. It definitely works - I pushed context to ~24% and it triggered compaction immediately (and kept compacting as I stayed over the limit). The model-specific overrides are doing their job. Thanks for adding this!

Thank you for implementing this! I have tested it with opencode 1.2.10.

Working config format:

{
  "compaction": {
    "token_threshold": 60000,
    "context_threshold": 0.45,
    "models": {
      "google/antigravity-gemini-3-flash": {
        "context_threshold": 0.30
      },
      "google/antigravity-gemini-3-pro-high": {
        "context_threshold": 0.30
      }
    }
  }
}

Note: The following keys from my previous config are not supported in this implementation:

• auto, prune, reserved, strategy, threshold, preserveRecentMessages, preserveSystemPrompt

These should be removed. The supported keys are only:

• token_threshold (global default)
• context_threshold (global default)
• models (per-model overrides with same keys)

Feature request: Would it be possible to add min_messages (minimum messages between compactions) to prevent continuous compaction loops? This would be especially useful for long-running sessions.

@WietRob how did you get this to work? I just updated my opencode to 1.2.12 and adjusted my opencode.json accordingly. Still, the auto compaction only kicks in at 75%.

// opencode.json
{
    "$schema": "https://opencode.ai/config.json",
    "compaction": {
        "token_threshold": 60000,
        "context_threshold": 0.45
    },
    "mcp": {
        ...
    }
}

nudge!

@fea17e86 Sorry for the confusion! I just tested this myself and found the issue.

Why it doesn't work for you:
The custom compaction thresholds (token_threshold, context_threshold) are part of PR #10123, which is still open and not yet merged. Version 1.2.12 does not include these features - it only has the default behavior (compaction at 75% context usage).

What I found when testing the PR branch:

• The PR code correctly reads compaction.token_threshold and compaction.context_threshold from opencode.json
• It also supports compaction.min_messages to prevent compaction loops
• Model-specific overrides via compaction.models work as documented

To use these features now:
You would need to build from the PR branch directly:

git clone https://github.com/anomalyco/opencode.git
cd opencode
git fetch origin pull/10123/head:pr-10123
git checkout pr-10123
# Follow the build instructions in the repo

Your config format is correct - it will work once this PR is merged and released.