Skip to content

opencode2.md

Latest commit

 

History

History
73 lines (56 loc) · 3.09 KB

opencode2.md

File metadata and controls

73 lines (56 loc) · 3.09 KB
Raw

OpenCode 2

The same package runs on OpenCode 1.x and OpenCode 2. Memory still lives in the plugin home (OpenCode data dir by default, or home / OPENCODE_CODEX_MEMORY_HOME) and uses the same options. Pin the version — OpenCode installs a plugin spec once and does not follow “latest”.

Install

In your OpenCode 2 config (opencode.jsonc or equivalent):

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": [
    { "package": "[email protected]", "options": { "min_rollout_idle_hours": 1 } },
  ],
}

OpenCode 1.x keeps the older "plugin": ["[email protected]"] form.

All V1 options apply unchanged: generate_memories, use_memories, dedicated_tools, disable_on_external_context, extract_model, consolidation_model, the numeric clamps, codex_interop, claude_import, and home.

What you get

On OpenCode 1.x the plugin still loads; there is no Memory sidebar there (that UI is OpenCode 2 only).

On OpenCode 2 the session sidebar grows a Memory section. /memory-status (also Show memory status in the command palette) opens the same snapshot memory_inspect uses: effective models, read/write settings, import status, retry eligibility, and warnings.

  • Status is global. Another window’s work shows up here.
  • The panel refreshes on its own while it is open. Opening it does not start extraction or consolidation.
  • Use memories / Learn from sessions / Learn from this session toggle injection and whether sessions are eligible.
  • Consolidate now extracts idle sessions and rebuilds the summary, skipping the usual 6-hour cooldown. Turn learning on first. It runs in the background.
  • If the OpenCode 2 service is missing or not this process, the panel shows Unavailable, not a stale Idle. Last success is only a clean consolidation timestamp; — means there isn’t one for that attempt.

Where it differs from 1.x

Same memories, not the same host. These are the limits that show up in use:

Background learning lists sessions from this machine’s registered local service. Sessions live in the shared OpenCode database, so an IDE serve --port 0 whose PID does not match service.json still uses a healthy loopback --service for the global list. A non-loopback PID mismatch is refused. If no service is registered, this process only extracts chats it has seen. Injection of an already-built summary always works.

Citation blocks can stay in the saved transcript so the sidebar can render them. They are stripped before the next model call. OpenCode 1.x removes them before the reply is stored.

Set extract_model and consolidation_model if you care which models run in the background. OpenCode 2 has no small_model. Unset extraction uses the session default; consolidation uses the configured model when present.

Helper work still cannot edit your repo. OpenCode 2 registers the memory agents in the project you have open, so those short-lived sessions are created there. Their file tools are allowlisted only under the memory folder.

Full 1.x install and configuration: README.