rz-ai-researcher

rz-agent-team

Versioned configuration for Riché's 11-agent OpenClaw team. The agents build and maintain the app portfolio: SIA, richezamor.com, and 6 prototypes (Recipe Remix, Ploppy, Blocade, Ascend, Trend Analyzer, AI Onboarding).

The authoritative design lives in Notion under 🤖 Agent Team. This repo is the deployable artifact — the markdown each OpenClaw instance reads at startup.

What's in here

.
├── README.md                  This file
├── TEAM.md                    Full roster — identical copy on every instance
├── USER.md                    Riché's working context + app registry pointers — identical copy on every instance
├── identities/                One IDENTITY.md per agent (11 total)
│   ├── conductor.md
│   ├── pm-lite.md
│   ├── researcher.md
│   ├── designer.md
│   ├── backend-eng.md
│   ├── data-eng.md
│   ├── ai-eng.md
│   ├── ui-eng.md
│   ├── qa-eng.md
│   ├── devops-eng.md
│   └── tech-writer.md
├── corpus/                    Knowledge corpus seeds for each role (Cowork research input)
├── connect.sh                 SSH helper for the Hostinger VPS
└── .env.local.example         Template for local connection vars

How OpenClaw consumes these files

Each of the 11 OpenClaw instances on the Hostinger VPS (/docker/openclaw-*/) loads three identity files at startup:

• TEAM.md → identical across all instances
• USER.md → identical across all instances
• IDENTITY.md → the role-specific file from identities/{role}.md, copied to the instance's working directory as IDENTITY.md

A future deployment script will sync these files from this repo to each /docker/openclaw-{role}/ directory on the VPS. Until that script exists, propagate by hand:

# Example for the conductor instance
scp identities/conductor.md [email protected]:/docker/openclaw-<conductor-instance>/data/IDENTITY.md
scp TEAM.md USER.md [email protected]:/docker/openclaw-<conductor-instance>/data/
docker compose -f /docker/openclaw-<conductor-instance>/docker-compose.yml restart

Required environment per instance

Every /docker/openclaw-*/.env on the VPS must contain:

ANTHROPIC_API_KEY=...   # Opus 4.7 — primary for Conductor + AI Eng, escalation for everyone else
OLLAMA_API_KEY=...      # Ollama Cloud — workhorse inference for the other 9 roles

Local Ollama (ollama-apvg container) on the VPS is reserved for embeddings only (nomic-embed-text model). Do not configure agents to use it for chat completion.

Operating model recap

• Session-based, human-triggered. No 24/7 autonomous operation.
• Every session targets exactly one app (app_id). Cross-app work splits into sequential sub-sessions.
• Two-layer orchestration: Linear = planning, Paperclip = execution + audit.
• Three-layer memory: identity files (static), Slack channels (live), Postgres agent_memory schema (persistent, partitioned by app_id).
• Observability: Langfuse traces every LLM call, tagged with app_id + agent role + session ID.

Full operating rules live in the Notion Operating Rules & Conventions page.

Setup prerequisites

1. Anthropic API key — Claude Opus 4.7 access. Required by every instance.
2. Ollama Cloud API key — sign up at ollama.com, generate a key. Required by the 9 workhorse instances.
3. Slack workspace — bot tokens for each of the 11 agent identities (@conductor, @pm, @researcher, @designer, @backend-eng, @data-eng, @ai-eng, @ui-eng, @qa-eng, @devops-eng, @tech-writer).
4. Linear API token — Conductor + PM-lite need write access; other agents read-only.
5. GitHub access — agents work directly on each app's repo (see Notion app registry for the 8 repo URLs).
6. Notion API token — agents read product specs and design pages.

Deferred work (not in this repo yet)

• Sync/deploy script (scripts/deploy-identities.sh)
• 6 additional openclaw instances to reach the design's target of 11 (currently 5 deployed)
• agent_memory Postgres schema migrations
• Slack channel + bot provisioning script
• Conductor's Linear ↔ Paperclip bridge implementation