DualView

DualView is an OpenClaw plugin that keeps untrusted tool output out of the trusted agent context. It replaces untrusted values with opaque symbols, tracks those symbols across tool calls and file writes, and lets the agent request isolated transformations through inspect_symbol when a task needs the underlying content.

This repository includes the DualView plugin, the OpenClaw source version used for this release, and the local dashboard.

Repository Layout

plugin/dualview/   DualView OpenClaw plugin
openclaw/          bundled OpenClaw source for this release
dashboard/         local DualView activity/session dashboard

Requirements

• Node.js 22.12 or newer
• npm
• Git

Install

npm install

Confirm that the bundled OpenClaw runtime starts:

npm run openclaw -- --version

OpenClaw Bot

See docs/setup.md for the full Slack quick start (creating the Slack app, required .env values, and both run paths).

There are two ways to run the bot. Both validate .env on startup and stop with a clear error if a required Slack token or LLM credential is missing, and both build the bundled OpenClaw automatically.

• Docker (recommended; isolated, persistent): ./scripts/setup-openclaw-bot.sh --channel slack
• Local (non-Docker; runs against ~/.openclaw): ./scripts/run-bot-local.sh

For a Slack/Telegram/Discord bot run, put credentials in .env or the shell environment, then run the setup script:

cp .env.example .env
# Fill Codex OAuth or another LLM auth method and one channel token set, then:
./scripts/setup-openclaw-bot.sh --channel slack --open-access

The script writes tmp-runs/bot/generated/openclaw-bot.json without channel tokens, starts Docker through scripts/run-bot.sh, and enables DualView for the configured channels. If channel token env vars are unset, it imports the default-account tokens from ~/.openclaw/openclaw.json into the process environment without writing them to the generated config file. The default model is openai-codex/gpt-5.3-codex-spark; override it with --model or DUALVIEW_BOT_MODEL. The inspect_symbol U-LLM fallback/logging model follows the same model unless DUALVIEW_BOT_INSPECT_MODEL is set.

Use the setup script for normal starts and restarts:

./scripts/setup-openclaw-bot.sh --channel slack --open-access

To run from a checked-in or local config file instead of CLI target/model flags, pass the config explicitly:

./scripts/setup-openclaw-bot.sh --config config/openclaw-bot-dualview.json

Secrets still come from .env, the shell environment, or ~/.openclaw/openclaw.json; do not write channel tokens into config files. The checked-in config/openclaw-bot-dualview.json is an example for a DualView-limited Slack channel. Copy it and replace the example channel ID for local deployments.

Bot config files may include runner-only settings under dualviewBot. These are read by scripts/setup-openclaw-bot.sh / scripts/run-bot.sh and stripped before the OpenClaw runtime config is written:

{
  "dualviewBot": {
    "port": 18801,
    "dashboard": {
      "port": 13457,
      "host": "dashboard.example.test",
      "timezone": "Asia/Seoul"
    }
  }
}

dualviewBot.port controls the gateway port. dualviewBot.dashboard.port controls the dashboard port and Slack symbol-link port. dualviewBot.dashboard.host controls the public host used in Slack links; leave host-specific values out of committed configs and set them in a local config or DUALVIEW_DASHBOARD_HOST. dualviewBot.dashboard.timezone controls dashboard timestamp display; omit it to use the system timezone.

Do not restart a Slack/Telegram/Discord bot with scripts/run-bot.sh directly unless the channel token environment variables are already exported in that shell. run-bot.sh only forwards existing env vars into Docker; it does not import default-account channel tokens from ~/.openclaw/openclaw.json.

By default, DualView targets every enabled external platform with entries such as slack:*. To keep the bot active broadly while disabling DualView in a specific channel, keep the broad target and pass a disable list:

./scripts/setup-openclaw-bot.sh --channel slack --open-access \
  --dualview-disable-targets slack:C0123456789

The target is the Slack channel ID, not the #channel-name. This is useful when the bot should respond in a channel such as #proj-adfi, but DualView should not inspect that channel. Disable entries take precedence over allowlist entries, so targetChannels: ["slack:*"] plus disabledTargetChannels: ["slack:C0123456789"] enables DualView for Slack except that one channel.

To enable DualView only for selected channels instead, pass an explicit target allowlist:

./scripts/setup-openclaw-bot.sh --channel slack --open-access \
  --dualview-targets slack:C0123456789

For a config-only dry run:

./scripts/setup-openclaw-bot.sh --channels slack,telegram --no-run

Dashboard

Start the dashboard:

npm run dashboard

Open http://127.0.0.1:3456/. The default dashboard is bot-only and redirects to /bot/. Override the bot log directory with DUALVIEW_BOT_DIR when viewing a different run directory.

Useful environment variables:

DASHBOARD_PORT=3456
DASHBOARD_PASSWORD=
DUALVIEW_BOT_DIR=/path/to/bot/log-sessions

Development

For normal bot runs, use the scripts above. They install and enable the DualView plugin as needed.

If you are developing the plugin against the bundled OpenClaw runtime manually:

npm install
npm install --prefix plugin/dualview
npm run openclaw -- plugins install ./plugin/dualview --link
npm run openclaw -- plugins enable dualview
npm run openclaw -- gateway run

--link makes OpenClaw load the plugin from this working tree instead of a copied plugin snapshot. That is useful for development because edits under plugin/dualview/ are picked up without reinstalling the plugin. If the repo is moved or deleted, the linked plugin path must be reinstalled.

DualView options are documented in plugin/dualview/openclaw.plugin.json. The most common settings are:

• targetSessionIds: session key fragments where DualView should run
• targetChannels: channel IDs or platform wildcards where DualView should run
• disabledTargetChannels: channel IDs or platform wildcards where DualView should not run, even if a broad target such as slack:* matches
• disabledTargetSessionIds: session key fragments where DualView should not run
• policyPath: path to a dualview-policy.yaml data trust policy
• fileTrackingEnabled: enable git-based trusted/untrusted file tracking

npm run typecheck

License

The original DualView code in this repository is licensed under the Business Source License 1.1 (BUSL-1.1). See LICENSE for the licensed-work scope and terms. On August 4, 2030, it will become available under the Apache License 2.0.

The bundled OpenClaw source, third-party dependencies, and files with separate license notices remain under their respective licenses.

Acknowledgements

Contributors:

• Juhee Kim (Seoul National University)
• Woohyuk Choi (Seoul National University)
• Taehyun Kang (Seoul National University)
• Youngmin Kim (Seoul National University)
• Byoungyoung Lee (Seoul National University)

This project was supported by LG Electronics.