decant

OpenCode plugin prototype for agent context maps.

decant turns a long agent chat into topics that can be shown at different fidelities: full text, summary, compressed, placeholder, or hidden. The point is to keep current work readable while old work stays recoverable.

It also experiments with provenance: /blame <file>:<line> tries to route from code back to the old agent session/message that explains why the line exists.

What's Here

• src/server-plugin.ts: server plugin, context tools, storage, blame lookup.
• src/tui-plugin.tsx: TUI sidebar and commands.
• test/: fast unit/regression tests.
• tools/: sandbox setup, validation, fixture, and artifact scripts.
• benchmarks/: benchmark harnesses and per-benchmark docs.
• artifacts/benchmark-runs/: curated benchmark evidence.
• fixtures/: small demo/validation data.

Install

npm install

Fast Dev Checks

npm run typecheck
npm test
npm run validate:blame-tui

Disposable Demo

npm run setup:test-env

That prints a launch script for a disposable OpenCode test repo. Run the script, then try:

/context
/blame src/auth/rate_limiter.ts:42

Model-Backed Validation

Use a GPT-family model for live validation and public evidence runs.

export DECANT_E2E_MODEL="<provider>/<model>"
npm run validate:sandbox
npm run validate:long-session
npm run evaluate:compaction

Manual Plugin Load

The plugin is not packaged. After local checks pass, link the entrypoints into a target project:

cd /path/to/target-project
mkdir -p .opencode/plugins
ln -s /path/to/decant/src/server-plugin.ts .opencode/plugins/context-map.ts
ln -s /path/to/decant/src/tui-plugin.tsx .opencode/plugins/context-map-tui.tsx

Add .opencode/tui.json if you want the TUI plugin:

{
  "plugin": ["./plugins/context-map-tui.tsx"]
}

Start OpenCode from the target project. Sideband maps are written to ~/.opencode/context-maps/<session-id>.json.

Benchmarks

Benchmark details live in each benchmarks/*/README.md. Raw runs stay ignored under benchmarks/*/runs/; curated public artifacts live under artifacts/benchmark-runs/.

Common commands:

npm run benchmark:memory-infra
npm run benchmark:provenance-qa
npm run benchmark:code-recall
npm run artifacts:export