opencode-agent-compaction

The agent compresses its own context — mid-work, automatically, non-blocking.

Agent-triggered context compaction for OpenCode: the agent calls the built-in compact_context tool between steps, so the standard OpenCode compaction runs mid-task and the agent continues where it left off. No custom compaction logic — just a thin wrapper around the built-in mechanism.

Quick Start

1. Copy plugins/compact-context.ts into ~/.config/opencode/plugins/ (global) or .opencode/plugins/ (project).
2. Restart OpenCode.

The agent now has the compact_context tool, which fires the compaction asynchronously — the running loop compacts between steps and the agent continues without pausing.

Agent Prompt Example

Add to your <agent>.md to let a long-running agent compact its own context as it goes (adjust the wording to your needs):

Before continuing to the next step, compact the current session's context:
    - Call the `compact_context` tool — compaction runs asynchronously between steps, work continues automatically.
    - After compaction — re-read the plan and reload your skills.

The optional reason parameter

The tool accepts a single optional parameter — reason, a short note on why compaction was requested. The agent may use it to record the intent:

compact_context(reason: "context too large, need to free tokens for the next phase")

Why

OpenCode compacts context automatically on overflow, but the agent has no control over when it happens. This plugin gives the agent that control — without ever blocking its own run.

Tested with OpenCode 1.17.18.