@madsoftwaredev/opencode-compaction-model

Summarizes local checkpoint compaction with a dedicated model and variant while the session keeps its selected working model.

opencode plugin add @madsoftwaredev/opencode-compaction-model

Choose the compaction model

Set options.model in your global ~/.config/opencode/opencode.json or your project's opencode.json (JSONC also works). If the CLI already added a string entry for this package, replace that entry with the object below, keeping your other plugins:

{
  "$schema": "https://opencode.ai/config.json",
  "plugins": [
    {
      "package": "@madsoftwaredev/opencode-compaction-model",
      "options": {
        "model": "deepseek/deepseek-flash#max"
      }
    }
  ]
}

Use provider/model#variant, with #variant optional. Choose an enabled model and supported reasoning variant from /models, using a provider connected in OpenCode. The example selects direct DeepSeek; you can select another provider and model the same way. Omitting #variant uses that model's default settings.

When model options are omitted, the default is opencode-go/deepseek-v4.1-flash#max, which requires an OpenCode Go connection. This default is overridable. The separate providerID, modelID, and variant options are also supported; a full model string takes precedence over them.

There is no default character cutoff. Optional maxTranscriptChars is a guard: exceeding it logs a fallback rather than silently deleting the middle of the history.

Why the Go adapter exists

In OpenCode 2.0.8–2.0.9, ctx.generate.text() makes a stateless request without x-opencode-session. OpenCode Go rejects that request. The plugin's Go adapter uses the existing OpenCode connection, catalog endpoint, model ID, and variant, and sends the originating session ID to the chat-completions endpoint. It does not store an API key or change the session's model. Other providers use the native ctx.generate.text() API.

Starting with 0.1.2, credential lookup checks the model's own provider integration first, then the integration ID reported by the provider, accepting only API-key credentials. OpenCode Go can report the Console integration, whose active connection may be OAuth rather than the Go API key. Earlier releases used only that reported integration, causing compaction to fall back even when a Go key was connected. The fix does not embed connection IDs, keys, or endpoint URLs; it uses the selected model and runtime provider metadata.

The Go adapter currently supports models using @opencode/ai/providers/openai-compatible (including the default DeepSeek model). Other Go protocols produce a diagnostic and fall back to normal compaction.

Check a compaction

Successful checkpoints carry this metadata:

{
  "compactionModel": "opencode-go/deepseek-v4.1-flash#max",
  "compactionPlugin": "compaction-model"
}

OpenCode still stamps the checkpoint's top-level model with the session's working model. Check metadata.compactionModel to identify the actual summarizer:

opencode api get '/api/session/SESSION_ID/message?type=compaction&order=desc&limit=1'

Successes and fallbacks are logged in ~/.local/share/opencode/log/compaction-model.log (or beneath $XDG_DATA_HOME). The log does not include transcripts, summaries, credentials, or Go response bodies.

Summary and failure behavior

• Sends the full supplied text/tool transcript, including prior checkpoints. Binary media is represented by a filename/type descriptor.
• Requires the checkpoint headings and retries an invalid summary once with the original transcript. Empty, truncated, or twice-invalid summaries are not installed.
• Go requests have a three-minute timeout and a 32,768-token output ceiling.
• On failure, logs the reason and lets OpenCode use the session's normal compaction.
• OpenCode owns automatic scheduling, recent-context retention, and checkpoint installation. The plugin handles requests delivered to the compaction hook; native provider checkpoint routes are outside this adapter's verification.

Verification

From the repository root:

bun test ./packages/compaction-model/index.test.js

The regression suite covers Go session routing, provider/integration mismatches and credential fallback, concurrent sessions, variant selection, long-history preservation, invalid/truncated results, and logged fallback. A live scratch session using Muse was manually compacted with DeepSeek and its checkpoint metadata verified.

References: V2 plugin hooks, V2 compaction.

License

MIT