opencode-dcp-dynamic-limits

OpenCode plugin for one specific annoyance: DCP should follow the context window of the model you are actually using.

I wrote this for my local llm-server setup, where OpenCode switches between GGUF models served by ik_llama.cpp / llama.cpp and normal cloud models. When the active model changes, DCP can be left with stale limits. That is when it starts pruning at the wrong time, or worse, keeps dropping dcp-system-reminder messages into the chat.

This plugin runs on real chat requests, figures out the active model context, and updates the DCP config before DCP needs it.

Install

npm install -g opencode-dcp-dynamic-limits

Add it after DCP in your OpenCode config:

{
  "plugin": [
    "@tarquinen/opencode-dcp@latest",
    "opencode-dcp-dynamic-limits"
  ]
}

DCP needs to load first. This plugin only adjusts DCP's config.

llm-server

The main target is llm-server:

llm-server model.gguf --port 8081 --ctx-size 262144

Point OpenCode at that local OpenAI-compatible endpoint as usual. The plugin will read the active provider/model and, for llama.cpp-style backends, use the running server's context instead of guessing from a static number.

It also works with plain llama.cpp-style servers, Ollama-style local providers, and cloud models from the OpenCode model manifest.

What It Changes

The plugin writes both common DCP config files when they exist:

~/.config/opencode/dcp.jsonc
~/.opencode/dcp.jsonc

It updates:

• activeModel
• minContextLimit
• maxContextLimit
• notification/nudge settings that keep reminders out of the chat

It ignores non-chat calls like the title agent, so a title-generation request cannot shrink your DCP limits.

Context Rules

For local llama.cpp-style servers:

• prefer runtime --ctx-size
• fall back to /v1/models metadata
• treat --parallel slot size as diagnostics, not the full context window

For cloud models:

• use OpenCode's model manifest context

The DCP limit is set to 85% of the detected context. Quiet mode sets minContextLimit equal to maxContextLimit, which avoids the repeated soft-warning zone while still letting DCP prune at the hard limit.

The quiet settings are:

{
  "pruneNotificationType": "toast",
  "compress": {
    "nudgeForce": "soft",
    "nudgeFrequency": 50,
    "iterationNudgeThreshold": 50
  }
}

Development

npm install
npm test
npm pack

License

MIT