@byterover/byterover

ByteRover context engine plugin for OpenClaw. Integrates the brv CLI as a context engine that curates conversation knowledge and retrieves relevant context for each prompt — giving your AI agent persistent, queryable memory.

Table of contents

• What it does
• Prerequisites
• Quick start
• Configuration
• How it works
• Development
• Project structure
• License

What it does

When you chat with an OpenClaw agent, the conversation is ephemeral — older messages get compacted or lost as the context window fills up. ByteRover changes that by:

1. Curating every turn — after each conversation turn, the plugin feeds the new messages to brv curate, which extracts and stores facts, decisions, technical details, and preferences worth remembering
2. Querying on demand — before each new prompt is sent to the LLM, the plugin runs brv query with the user's message to retrieve curated knowledge relevant to the current request
3. Injecting context — retrieved knowledge is appended to the system prompt so the LLM has the right context without the user needing to repeat themselves

The result: your agent remembers what matters, forgets what doesn't, and retrieves context that's actually relevant to what you're asking about right now.

Prerequisites

• OpenClaw with plugin context engine support
• Node.js 22+
• brv CLI installed and available on your PATH (or provide a custom path via config). The brv path depends on how you installed it:

Quick start

1. Install the plugin

openclaw plugins install @byterover/byterover

For local development, link your working copy instead:

openclaw plugins install --link /path/to/brv-openclaw-plugin

2. Configure the context engine slot

openclaw config set plugins.slots.contextEngine byterover

3. Set plugin options

Point the plugin to your brv binary and project directory:

openclaw config set plugins.entries.byterover.config.brvPath /path/to/brv
openclaw config set plugins.entries.byterover.config.cwd /path/to/your/project

4. Verify

openclaw plugins list

You should see byterover listed and enabled. Restart OpenClaw, then start a conversation — you'll see [byterover] Plugin loaded in the gateway logs.

5. Uninstall (if needed)

openclaw plugins uninstall byterover
openclaw config set plugins.slots.contextEngine ""

Configuration

ByteRover is configured through plugins.entries.byterover.config in your OpenClaw config file (~/.openclaw/openclaw.json):

{
  "plugins": {
    "slots": {
      "contextEngine": "byterover"
    },
    "entries": {
      "byterover": {
        "enabled": true,
        "config": {
          "brvPath": "/usr/local/bin/brv",
          "cwd": "/path/to/your/project",
          "queryTimeoutMs": 12000,
          "curateTimeoutMs": 60000
        }
      }
    }
  }
}

Options

How it works

ByteRover hooks into three points in the OpenClaw context engine lifecycle:

afterTurn — curate conversation knowledge

After each conversation turn completes, the plugin:

1. Extracts new messages from the turn (skipping pre-prompt messages)
2. Strips OpenClaw metadata (sender info, timestamps, tool results) to get clean text
3. Serializes messages with sender attribution
4. Sends the text to brv curate --detach for asynchronous knowledge extraction

Curation runs in detached mode — the brv daemon queues the work and the CLI returns immediately, so it never blocks the conversation.

assemble — retrieve relevant context

Before each prompt is sent to the LLM, the plugin:

1. Takes the current user message (or falls back to scanning message history)
2. Strips metadata and skips trivially short queries (< 5 chars)
3. Runs brv query with a 10-second deadline
4. Wraps the result in a <byterover-context> block and injects it as a system prompt addition

If the query times out or fails, the conversation proceeds without context — it's always best-effort.

compact — delegated to runtime

ByteRover does not own compaction. The plugin sets ownsCompaction: false, so OpenClaw's built-in sliding-window compaction handles context window management as usual.

ingest — no-op

Ingestion is handled by afterTurn in batch (all new messages from the turn at once), so the per-message ingest hook is a no-op.

Development

# Install dependencies
npm install

# Type check
npx tsc --noEmit

# Run tests
npx vitest run --dir test

# Link for local testing with OpenClaw
openclaw plugins install --link .
openclaw config set plugins.slots.contextEngine byterover

Testing locally

1. Initialize a brv project: cd /your/project && brv init
2. Link the plugin and configure as shown in Quick start
3. Restart OpenClaw
4. Send a few messages — check gateway logs for:

• [byterover] Plugin loaded — plugin registered
• afterTurn curating N new messages — curation running
• assemble injecting systemPromptAddition — context being retrieved and injected

Project structure

index.ts                    # Plugin entry point and registration
openclaw.plugin.json        # Plugin manifest (id, kind, config schema)
src/
  context-engine.ts         # ByteRoverContextEngine — implements ContextEngine
  brv-process.ts            # brv CLI spawning (query, curate) with timeout/abort
  message-utils.ts          # Metadata stripping and message text extraction
  types.ts                  # Standalone type definitions (structurally compatible with openclaw/plugin-sdk)

License

Elastic License 2.0 (ELv2)