research: Phase 0 reports 2 and 3 — gateway internals + live state audit

This commit is contained in:
Caret
2026-04-06 12:50:22 +00:00
parent f3296d136a
commit d36fa6538a
2 changed files with 295 additions and 0 deletions

View File

@@ -0,0 +1,163 @@
# Research 02 — openclaw gateway internals
**Subagent:** `ae5ca38f70b1e9626` (Explore)
**Completed:** 2026-04-06 12:50 UTC
## Gateway API surface
WebSocket-first RPC at `ws://localhost:18789/`, with HTTP fallback routes.
### HTTP endpoints
| Method | Path | Purpose |
|--------|-------------------------------------|-------------------------------------------------------------------------|
| POST | `/hooks/{hookPath}/wake` | Trigger heartbeat or immediate agent wake. Body: `{text, mode}`. |
| POST | `/hooks/{hookPath}/agent` | Spawn isolated agent session. Body: `{agentId, sessionKey, message, channel, to, deliver, model, thinking, timeoutSeconds}`. Returns `{ok, runId}`. Idempotency: 60s dedup by `Authorization + X-Idempotency-Key`. |
| POST | `/tools/invoke` | Call a tool directly. Body: `{tool, action, args, sessionKey, dryRun}`. |
| GET | `/health` / `/healthz` / `/ready` | Liveness / readiness probes. |
| GET | `/` and `/app/*` | Built-in web control UI (the SPA we saw when probing earlier). |
| Plugin-registered routes | Custom plugin HTTP endpoints; auth enforced per plugin's `requiresAuth`.|
### Authentication
- `Authorization: Bearer <token>` OR `X-OpenClaw-Token: <token>` header
- Token sources: `gateway.auth.token` in config, `OPENCLAW_GATEWAY_TOKEN` env var, device token at `~/.openclaw/credentials/device-token`
- WebSocket auth: passed in URL query `?token=...` or connect frame
### RPC method RBAC scopes
- READ: `health`, `channels.status`, `sessions.list`, `cron.list`, `node.list`, ...
- WRITE: `send`, `agent`, `agent.wait`, `wake`, `node.invoke`, ...
- ADMIN: `config.set`, `agents.create`, `cron.add`, `sessions.reset`, ...
- APPROVALS, PAIRING: narrower scoped methods.
## Session spawn recipe
### The primary spawn path
```
Client RPC request → gateway dispatch → agentHandlers.agent() → agentCommandFromIngress() → in-process task
```
Not a child process. Sessions run as in-process tasks under the gateway. Each session's message history lives in `~/.openclaw/sessions/*.jsonl`.
### Agent identity & tool allowlist resolution at spawn
1. Resolve agent ID from `params.agentId` or `agents.defaults.id`.
2. Resolve tool allowlist: first match wins among `agents[id].tools.allow/deny``agents[id].toolProfile``agents.defaults.tools.*` → subagent role restrictions.
3. Hard-deny list always wins (`exec.approval.*`, `node_invoke_system_run`, etc.).
4. Runtime context: `runtime="subagent"` (sandboxed) or `"acp"` (host access).
5. Workspace and session store selected from agent's config.
### Subagent / ACP spawn (for nesting)
```typescript
const result = await spawn({
task: "Analyze the attached image",
mode: "run" | "session",
thread: true,
agentId: "analyzer"
});
// Returns { status, childSessionKey: "subagent:uuid", runId }
```
Sessions prefixed `subagent:*` run in a sandbox (gVisor or Docker container). `acp:*` runs on host under parent's cwd. Parent sees subagent output but can't reach into its filesystem.
## Cron / heartbeat mechanism
**It's not a crontab. It's an in-process scheduler built into the gateway.**
### Heartbeat loop
1. At gateway boot, `startHeartbeatRunner()` in `src/infra/heartbeat-runner.ts` starts.
2. For each agent where `agents[id].heartbeat.enabled == true`:
- Parse `heartbeat.every` interval
- Calculate next-due time
- Set a timer (internally a `setInterval` that checks wall clock every ~10s)
3. When timer fires:
- Read `memory/heartbeat-state.json` (for dedup / avoid double-fires)
- Read pending `memory/system-events/` (queued by cron jobs, exec completions, etc.)
- Build a prompt from heartbeat config + pending events
- Spawn agent with `extraSystemPrompt` = heartbeat prompt
- Agent responds (may be empty)
- Update heartbeat state file
### Cron service (parallel to heartbeat)
- Class: `CronService` in `src/cron/service.ts`
- Config: `cron.jobs[].schedule` (cron expression)
- State: `~/.openclaw/memory/cron/store.json` with `{id, schedule, agentId, prompt, lastRunMs, nextDueMs}`
- Run logs: `~/.openclaw/memory/cron/runs/`
- Can enqueue `system-events/*.json` that heartbeat picks up next cycle.
### Ad hoc triggers
- `openclaw wake --now` fires heartbeat immediately
- `openclaw cron run <id> --force` fires a cron job immediately
- `openclaw system-event "text"` queues an event for next heartbeat
## Plugin discovery and wiring
### Loader
`src/plugins/loader.ts``loadOpenClawPlugins()`:
1. Scan `~/.openclaw/plugins/` directory
2. Read each plugin's manifest (plugin.yaml or package.json exports)
3. Dynamic-import plugin module via jiti
4. Initialize `PluginRuntime` with sandbox context, gateway request handler, scoped filesystem access
5. Register plugin's hooks (lifecycle events) and gateway methods (HTTP/RPC)
### Example: Telegram plugin
- Starts a polling loop calling Telegram Bot API `getUpdates()`
- For each incoming message, calls `dispatchGatewayMethod("agent", {...})` to spawn a Claude session
- Claude's response routed back via plugin's send handler
## Replacement difficulty matrix
| Component | Difficulty | Notes |
|--------------------------------------------|-----------|----------------------------------------------------------------|
| Session storage (JSONL messages) | Easy | Simple file format, adopt as-is |
| Heartbeat scheduler | Medium | Timer logic easy; state/dedup is the work |
| Cron service | Medium | Schedule parsing + state persistence |
| Hook API (POST /hooks) | Easy | Stateless request/response |
| RPC / WebSocket protocol | Hard | Custom protocol with dedup, framing, RBAC |
| Tool policy and allowlist resolution | Medium | Glob pattern + inheritance hierarchy |
| Plugin system | Hard | Dynamic loading, sandboxed runtime contexts |
| Subagent / ACP spawn | Hard | Nesting, thread binding, runtime isolation |
| Delivery system (Telegram, Slack, etc.) | Hard | Multi-channel abstraction; tightly coupled |
| Control UI | Medium | React SPA; can be replaced if protocol stays compatible |
| Authentication and RBAC | Medium | Token validation + scope checks |
## Don't reinvent this
1. **Session transcript storage** (`src/config/sessions/`) — JSONL with dedup, compression, archival. Adopt.
2. **Plugin SDK** (`src/plugin-sdk/`) — type-safe hook runners, tool registration. Many plugins depend on it.
3. **Tool policy resolution** (`src/agents/tool-policy*.ts`) — battle-tested glob + inheritance. 2-3 weeks to replace.
4. **Delivery system** (`src/infra/outbound/`) — routes to Telegram/Slack/Discord/WhatsApp with retries and dedup. Very tightly coupled.
5. **Exec approvals** (`src/infra/exec-approvals-*`) — human-in-the-loop for sensitive ops. Keep if you plan approvals.
6. **Hot-reload config** (`src/gateway/config-reload.ts`) — atomic updates with broadcasts.
## Migration path summary
To replace openclaw's orchestration while keeping agents and tools:
1. Adopt existing session storage (or thin DB adapter)
2. Keep plugin system — at minimum the hook-runner pattern for startup/shutdown
3. Reimplement heartbeat scheduler as a background job
4. Reimplement cron service with same semantics
5. Build your own HTTP/RPC gateway, keeping `/tools/invoke` signature for compatibility
6. Map hook API to your agent spawn endpoint
7. Reimplement tool policy resolution using your config schema
8. Adopt delivery system or build equivalent (biggest lift)
**Estimated effort:** 4-8 weeks competent team, assuming Claude SDK agent harness is mostly intact and session/tool abstractions reused.
## Caret's conclusion
Full orchestration replacement is a 4-8 week project. That's NOT what I want.
**What I DO want is much smaller**: the specific slice that handles Gitea webhook events → policy enforcement → optional agent wake-up. That's a ~600-800 line bun listener, not a whole orchestrator. Everything else (session storage, plugin SDK, delivery system, tool policy) I keep depending on openclaw for, or reuse Claude Code's native primitives (Channels plugins, CronCreate, hooks).
The research confirms the right shape: build a **minimal webhook listener + event router + script fan-out** that can run standalone, and wire it into Claude Code's native Channels mechanism for the judgment wake-ups. Don't try to replicate the whole orchestrator.