Skip to content

OpenCode

OpenCode (anomalyco/opencode) has first-class compaction, including a model setting that is the direct analogue of Zed's agent.compaction_model. Two routes work: native client-side compaction, or gateway-side through LiteLLM.

Declare RustySqueezer through an OpenAI-compatible provider, then name it as the compaction model:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "litellm": {
      "npm": "@ai-sdk/openai-compatible",
      "baseURL": "http://localhost:4000/v1",
      "models": {
        "rustysqueezer": {
          "name": "RustySqueezer",
          "limit": { "context": 1000000, "output": 4096 }
        }
      }
    }
  },
  "agents": {
    "compaction": { "model": "litellm/rustysqueezer" }
  },
  "compaction": {
    "auto": true,
    "keep": { "tokens": 15000 }
  }
}

If the model cannot be resolved

OpenCode falls back to the session's own model for the summary. So a typo in agents.compaction.model degrades silently rather than failing - check the RustySqueezer log to confirm requests are actually arriving.

What the knobs do

Key Default Behaviour
compaction.auto true Compact automatically near the context limit, and recover once when the provider rejects a request as too long
compaction.keep.tokens 15000 Roughly how much recent conversation is kept verbatim next to the summary
compaction.buffer 15% of the limit, minimum 16k Tokens kept free below the model's limit; larger means compaction starts earlier
agents.compaction.model session model Which model writes the summary

Route B - gateway side

Point OpenCode at LiteLLM and let the auto-router's context_compaction do the work, as described in the LiteLLM guide. This is the option when several clients share one gateway.

Provider-native compaction does not apply

OpenCode also supports per-provider settings.compaction: { "type": "native" }, where the provider compacts the conversation on its own side. That is OpenAI Responses-API specific and has no effect on RustySqueezer.

Improve OpenCode summaries

OpenCode sends its own summarisation instructions with the thread. RustySqueezer treats recognised instruction text as the request, not as thread content - add OpenCode's wording to instruction_patterns via CONFIG_PATH (see Configuration) so it is not counted as user input.