Aura is a reliability & memory control plane for a production Slack agent.
It sits beside OpenClaw and turns every user turn into a durable Temporal workflow: intake routing, process memory, evidence gates, idempotent Slack delivery, multi-key LLM orchestration, and a routed galaxy web-research stack.
Built by Hyper-AI-Lab · Homepage: hyperailab.com
Chat-agent stacks are great at tools and models — and terrible at ops truth:
- Slack replies race with native gateway delivery
- Sessions forget process context across turns
- Rate limits stall the whole agent with no fair key rotation
- “Done” is whatever the model claimed, not what evidence allows
- Canary restarts can kill mid-flight user work if remediation is too aggressive
Aura (the Reliability & Memory Plane, RMP) is the sidecar that owns those guarantees while OpenClaw stays the execution engine.
| Capability | What Aura provides |
|---|---|
| Durable task intake | 3-layer funnel (fast path → vector gate → LLM classify) with off / shadow / enforce modes |
| Workflow control plane | Temporal GenericTask / CatalogTask workflows, child steps, reconciler + janitor |
| Process memory | Process-scoped recall + promotion; Qdrant vectors (nv-embed-v1) |
| Slack ownership | OpenClaw plugin routes DMs to RMP; RMP posts the final reply (no double-send / no native fallback) |
| LLM orchestration | Balanced NVIDIA key rotation, concurrency caps, fast idle rotate (~5s), usage ledger |
| Galaxy web stack | Brave + LangSearch search; Jina Reader; Crawl4AI / Scrapling / Crawlee / ScrapeGraphAI; OpenClaw browser + browser-use + Obscura CDP |
| Web capability routing | Intake analyzer picks search / fetch / crawl / extract / interact and injects a tool brief |
| Production gates | Readiness API, hourly canaries with soft-fail deferral (no worker restart while user tasks run), orphan-reply Slack recovery |
flowchart TD
User["Slack_user_DM"] --> OC["OpenClaw_gateway"]
OC --> Plugin["rmp_adapter_claim"]
Plugin -->|"POST_/tasks"| API["RMP_FastAPI"]
API --> Intake["3_layer_intake"]
Intake --> WebCap["WebCapabilityAnalyzer"]
Intake --> Mode{"execution_mode"}
Mode -->|conversational_or_structured| Generic["GenericTaskWorkflow"]
Mode -->|interact_gated| Catalog["CatalogTask_browser_automation"]
WebCap -.->|preferred_tools_brief| Generic
WebCap -.->|preferred_tools_brief| Catalog
Generic --> Worker["rmp_worker"]
Catalog --> Worker
Worker -->|"hooks/agent_rmp_task"| OC2["OpenClaw_execution"]
OC2 --> Tools["Tools"]
Tools --> Native["web_search_web_fetch_browser"]
Tools --> AuraWeb["aura_web_plugin"]
AuraWeb --> LangSearch["LangSearch"]
AuraWeb --> Jina["Jina_Reader"]
AuraWeb --> Stack["web_stack_:8791"]
Stack --> Crawl4AI
Stack --> Scrapling
Stack --> Crawlee
Stack --> ScrapeGraph
Stack --> BrowserUse["browser_use"]
Stack --> Obscura["Obscura_CDP_:9222"]
Worker --> PG["PostgreSQL"]
Worker --> Qdrant["Qdrant"]
Worker -->|"notify_slack_user"| Slack["Slack_DM_idempotent"]
Canary["hourly_health_canary"] --> Sentinel["canary_sentinel"]
Sentinel -->|"soft_timeout_+_active_users"| Defer["defer_worker_restart"]
Sentinel -->|"hard_stale_or_code_sync"| Restart["restart_rmp_api_worker"]
Reconciler["reconciler"] -->|"orphan_OpenClaw_reply"| Slack
| Layer | Role |
|---|---|
| OpenClaw | Slack socket, LLM/tools, isolated rmp_task_* sessions (execution only — not Slack delivery owner) |
RMP (app/) |
API, workflows, intake, memory, quota broker, evidence, canary sentinel, reconciler |
Plugin (plugins/rmp_adapter) |
Intercepts Slack → creates RMP tasks; suppresses native double-posts (fail closed) |
Web (plugins/aura_web, plugins/langsearch, web-stack/) |
Multi-backend search/fetch/crawl/extract/browser tools + localhost FastAPI backends |
Binding rules: every Slack DM goes through RMP; MiniMax M3 is the primary chat model; LLM idle silence fails fast (~5s) and rotates NVIDIA keys.
Deep dive: ARCHITECTURE.md · Runbooks: docs/runbooks/
aura/
├── app/ # FastAPI + Temporal + memory + intake + web routing
├── plugins/
│ ├── rmp_adapter/ # Slack claim → RMP tasks
│ ├── aura_web/ # Galaxy web tools (Jina, Crawl4AI, …)
│ └── langsearch/ # LangSearch web_search provider + API key holder
├── web-stack/ # Local FastAPI backends + Obscura compose/systemd
├── ops/ # Canaries, backup, janitor, patch verify
├── tests/ # Pytest suite
├── docs/ # Runbooks, history, architecture assets
├── patch_openclaw.sh # Re-apply dist patches after OpenClaw upgrades
├── settings.example.json # Config template (no secrets)
├── worker.py # Temporal worker entrypoint
└── ARCHITECTURE.md # Full system design
- Linux host (or VM) with Docker optional for Qdrant / Obscura / observability
- Python 3.12+, Node.js ≥ 22.23 (OpenClaw engines)
- PostgreSQL, Temporal, OpenClaw gateway
- NVIDIA NIM (or compatible) API keys for chat + embeddings
- Optional: Brave + LangSearch API keys; Obscura image
h4ckf0r0day/obscura
git clone https://github.com/Hyper-AI-Lab/aura.git
cd aura
python3 -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp settings.example.json settings.json
# set api_key, production.slack_owner_user_id, vector/qdrant, task_registry.intake_mode
# Link plugins into your OpenClaw plugins dir (rmp_adapter, aura_web, langsearch), then:
bash patch_openclaw.sh
bash ops/verify_openclaw_patch.sh
# Optional galaxy web backends + Obscura CDP
# systemctl enable --now aura-web-backends aura-obscura
# Start API + worker (systemd units or process manager of your choice)
# then:
make production-checkmake production-check # health + OpenClaw patch verify + intake canaries
make canary # manual E2E canary task
pytest -q # unit/integration tests
curl -s http://127.0.0.1:8791/health # web-stack backends (if enabled)- Never commit
settings.json,.env, auth profiles, ordata/. - Example config:
settings.example.json. - LangSearch / Jina keys live in OpenClaw
plugins.entries.*(not this repo). - Obscura remote mode:
OBSCURA_CDP_URL=http://127.0.0.1:9222(Hermes-compatible). - After every
npm install -g openclaw, re-runpatch_openclaw.sh(hook persistence, Slack suppress, allowUnsafe passthrough, ~5s LLM idle). - Model stack (typical): MiniMax M3 primary → DeepSeek V4 Flash → GLM-5.2; intake/subagents on DeepSeek Flash.
- Health canary soft failures (
timeout/failed) defer worker restart while user tasks are active; reconciler can recover finished OpenClaw replies to Slack if delivery was interrupted.
This repository is a production-shaped public snapshot of Aura’s RMP control plane. Paths and host assumptions in older docs may reflect the original single-VPS deployment; adapt ports, systemd units, and secrets to your environment.
See CONTRIBUTING.md. Issues and PRs welcome for docs, tests, and portable packaging improvements.
MIT © Hyper-AI-Lab