openclaw-extensions

An OpenClaw extension repository by IYENTeam.

OpenClaw is an extensible agent runtime: it exposes a plugin contract (registerContextEngine, afterTurn, assemble, compact, PluginRuntime.subagent.*, …) and loads packages whose package.json declares an openclaw.extensions entry. This repository is a workspace of such extensions.

It currently ships a single extension — @iyen/assei — an external-verifier context engine: after every agent turn it asks an independent model "is more work needed?" and, if yes, auto-continues the same OpenClaw session until that verifier says DONE or BLOCKED.

Status: prototype. The plugin contract follows OpenClaw's published extension API; APIs may shift with upstream OpenClaw releases.

---

What this extension plugs into (OpenClaw vocabulary)

OpenClaw concept	How @iyen/assei uses it
Plugin manifest (openclaw.plugin.json)	declares plugin id assei and configSchema (model required)
Plugin entry (package.json → openclaw.extensions)	exports register(api) from index.mjs
Plugin slot (plugins.slots.contextEngine)	binds assei as the active context engine
Allow-list (plugins.allow)	must include "assei" for the runtime to load it
Hook: afterTurn(params)	runs the verifier and decides CONTINUE / DONE / BLOCKED
Runtime: PluginRuntime.subagent.*	invokes the verifier model — OpenClaw owns provider/auth/billing
Runtime: openclaw agent --session-id	spawned (detached) for the auto-continuation when verdict is CONTINUE

If you are not yet familiar with OpenClaw's plugin model, read the upstream extension docs first; the rest of this README assumes it.

---

For agents installing on behalf of a user

If you are an OpenClaw / Codex / Claude Code agent reading this, follow the rules below to install Assei into the user's OpenClaw without ambiguity. Assei is the only extension in this repo — there is no "pick one"; just install it if the user wants the autonomous-continuation loop on top of OpenClaw.

Step 1 — confirm intent

Install @iyen/assei if (and only if) the user wants any of the following layered on top of their OpenClaw runtime:

• "keep working / continue automatically / run until done / autonomous loop"
• "use a different / cheaper / external model only for the keep-going decision"
• "verify each turn against a second model before stopping"

If the user wants smart context strategy (compact / rewind / clear / spawn subagent), this repo does not provide an OpenClaw extension for that anymore. Tell the user and stop — do not install Assei as a substitute.

Step 2 — pick a verifier model before installing

Assei is by definition an external verifier. It must be pinned to a model that is independent of the main OpenClaw agent. If the user has not already chosen one, ask:

"Assei needs an external verifier model — independent of your main OpenClaw agent — to decide CONTINUE/DONE/BLOCKED each turn. Pick a cheap/fast model OpenClaw can already invoke, e.g. openai/gpt-5-mini, anthropic/claude-haiku-4, apiclient/glm-5.1. Which one should I use?"

Do not silently default to whatever the main agent uses. The plugin will refuse to run without assei.model set, and a validator_failed entry will appear in the audit log every turn until you fix it.

Step 3 — install the extension into OpenClaw

# 1. Clone or pull this extension repo somewhere stable on the user's machine
REPO=$HOME/.openclaw/extensions/openclaw-extensions
git clone https://github.com/IYENTeam/openclaw-extensions.git "$REPO" 2>/dev/null \
  || git -C "$REPO" pull --ff-only

# 2. Register the extension with OpenClaw (link mode)
openclaw plugins install --link "$REPO/packages/assei"

openclaw plugins install --link reads packages/assei/package.json, discovers the openclaw.extensions entry, and wires the plugin into the OpenClaw plugin loader without copying files.

Step 4 — bind the extension to the contextEngine slot

OpenClaw needs to know which extension to use as its context engine. Patch ~/.openclaw/openclaw.json to set the slot, the allow-list, and the required verifier model. Make a backup first:

cp ~/.openclaw/openclaw.json ~/.openclaw/openclaw.json.bak.$(date +%s)

Merge this block into the existing plugins object (do not replace the whole file):

{
  "plugins": {
    "slots": { "contextEngine": "assei" },
    "allow": ["assei"],
    "entries": {
      "assei": {
        "enabled": true,
        "config": {
          // REQUIRED — Assei is an EXTERNAL verifier and must be pinned
          // to a model that is independent of the main OpenClaw agent.
          // Replace with the model id the user picked in Step 2.
          "model": "<provider>/<verifier-model-id>"
        }
      }
    }
  }
}

Step 5 — verify the extension loaded

openclaw plugins list | grep assei
# Expected: a row with Status=loaded.

openclaw plugins doctor
# Expected: "No plugin issues detected."

If both pass, the extension is registered with OpenClaw. The user must restart the gateway (openclaw gateway restart if running as a service, or simply re-run their agent) for the new context-engine slot binding to take effect.

Hard rules for agents

• Never install Assei without an explicit assei.model set to a model different from the main OpenClaw agent. The plugin refuses to run otherwise and writes validator_failed: assei.model is required to the audit log every turn. If the user has not chosen a verifier model, ask first — do not pick one silently.
• Never modify any other key in ~/.openclaw/openclaw.json (channels, agents, gateway.auth, etc.). The user owns those.
• Never install Assei when the user only runs openclaw agent --local. Assei requires a real OpenClaw gateway runtime; --local does not expose PluginRuntime.subagent, so the verifier silently no-ops.
• Always back up ~/.openclaw/openclaw.json before patching it.

---

Extensions in this repo

Extension	Plugin id	OpenClaw slot	One-line description
@iyen/assei	assei	contextEngine	After every OpenClaw turn, ask an independent verifier model "is more work needed?" — if yes, auto-continue the same session via openclaw agent --session-id.

See packages/assei/README.md for the full configuration reference, OpenClaw runtime requirements, audit-log format, and programmatic API.

---

Repository layout

openclaw-extensions/
├── package.json                         # workspace root (pnpm)
├── pnpm-workspace.yaml
└── packages/
    └── assei/                            # OpenClaw extension package
        ├── package.json                  # @iyen/assei (declares openclaw.extensions)
        ├── openclaw.plugin.json          # OpenClaw plugin manifest (id, configSchema)
        ├── index.mjs                     # plugin entry — register(api) → registerContextEngine
        ├── assei-loop.mjs                # verifier loop core (uses runtime.subagent)
        ├── test-assei-loop.mjs           # unit tests (mocked subagent)
        ├── test-plugin.mjs               # plugin integration tests against a fake api
        ├── test-e2e-gateway.sh           # real-OpenClaw-gateway E2E
        └── README.md

Development

pnpm install
pnpm -r run test                                       # unit tests (mocked subagent)
ASSEI_E2E_MODEL=apiclient/glm-5.1 \
  pnpm --filter @iyen/assei test:e2e                   # real OpenClaw gateway E2E
                                                       # (mutates ~/.openclaw, restores on exit)

License

MIT — see LICENSE.