Setup Client Workspace (TNCG)

Mise en place d'un workspace agent dédié pour un nouveau client TNCG. Utiliser quand Erwan demande de configurer un nouvel agent client (workspace, fichiers...

Security Scan

The skill's procedures match a legitimate client-onboarding purpose, but the runtime instructions reference and instruct placement of multiple sensitive secrets, system-level changes (gateway restart, socat bridges), and sandbox tool allowances without declaring those env/config requirements — the power requested is plausible but larger than what the skill metadata advertises.

Purpose & Capability

The name/description align with the instructions: creating a workspace, generating SOUL/IDENTITY/USER/MEMORY/TOOLS files, configuring openclaw.json, pairing WhatsApp and enabling a sandbox are coherent tasks for a setup/orchestrator skill. The requested actions (editing workspace files, creating agent entries, WhatsApp pairing) are expected for onboarding.

Instruction Scope

The SKILL.md and referenced docs instruct reading/writing system paths (e.g. ~/.openclaw, /home/ubuntu/.openclaw/agents/...), editing openclaw.json, restarting the gateway, running docker/systemctl checks, creating socat bridges and copying/ installing other skills. They also reference environment variables and secrets (HOOKS_TOKEN, GATEWAY_HOST, GATEWAY_PORT, NANGO_PROD_SECRET_KEY, TAVILY_API_KEY, JINA_API_KEY, per-agent OPENAI_API_KEY, etc.) that are not declared in metadata. Because the skill is instruction-only, these instructions are the security surface — they ask an operator/agent to touch system-wide config and sensitive credentials, which is within the setup scope but must be made explicit to downstream operators.

Install Mechanism

This is an instruction-only skill with no install spec and no bundled code to execute. It references running local commands like 'clawhub install' and copying skills from other workspaces, but there are no external download URLs or archive extraction steps in the skill files. That lowers tooling-install risk, though the 'clawhub install' step will pull external packages at runtime (not described in metadata).

Credentials

The documentation requires multiple sensitive keys and secrets (per-agent OPENAI_API_KEY to be stored in auth-profiles.json and in container env, NANGO_PROD_SECRET_KEY, TAVILY_API_KEY, JINA_API_KEY, HOOKS_TOKEN, gateway host/port, etc.). These are relevant to the onboarding workflow, but the skill metadata declares no required env/config items. Storing API keys into agent auth JSON files (written to disk) and propagating keys into sandbox envs is justified for Whisper/SDK access, but is sensitive and should be explicit in the skill metadata and operator consent. The number and breadth of secrets is significant for a single onboarding playbook.

Persistence & Privilege

The skill's procedures modify global/system configuration (openclaw.json), add agents to inter-agent allow lists, copy other skills, and restart the gateway — actions that change system-wide state. This is expected for a setup/orchestrator but represents high privilege. The skill metadata does not request 'always: true' and is user-invocable, which is appropriate; nevertheless, the operator must be aware these steps will affect other agents and services.

What to consider before installing

This skill is plausibly a legitimate onboarding playbook, but it performs system-level configuration and asks you to store and propagate many secrets without declaring them. Before installing or running it: 1) Verify you trust the source (the files will edit ~/.openclaw and openclaw.json and restart the gateway). 2) Back up openclaw.json and any system config and confirm who is authorized to make these changes. 3) Confirm where API keys will be stored — the procedure writes per-agent OpenAI keys into auth-profiles.json and into sandbox envs; ensure keys are generated per-agent and have least privilege. 4) Audit any 'clawhub install' packages and copied skills (self-improving-agent, tavily-search, escalation) before running them. 5) Pin sandbox images and avoid wildcard allowSendTo = ["*"] unless you accept sending messages to arbitrary numbers. 6) Make the required secrets and env vars explicit in your deployment checklist (HOOKS_TOKEN, GATEWAY_HOST/PORT, NANGO_PROD_SECRET_KEY, TAVILY_API_KEY, JINA_API_KEY, per-agent OPENAI_API_KEY). 7) If an autonomous agent will execute these instructions, restrict that agent's rights and require human approval for the actual config changes. If you want a stricter assessment, provide logs of any 'clawhub install' operations, the openclaw.json target before/after, and the exact commands you expect the agent to run so I can check for inconsistencies or hidden endpoints.

Setup Client Workspace (Orchestrateur)

Ce skill est l'orchestrateur. Il reste court par design. Les détails opérationnels sont dans references/.

Règle d'exécution

Quand Erwan demande de créer/configurer un agent client :

1. Lire references/workflow.md (vue d'ensemble + ordre des étapes)
2. Lire references/templates.md pour générer les fichiers workspace
3. Lire references/openclaw-config.md pour la config JSON, WhatsApp, bindings, sandbox
4. Lire references/security-checklist.md avant validation finale

Ne pas improviser hors de ces procédures sans le signaler explicitement.

Minimum à collecter avant setup

• Nom du client / entreprise
• Rôle de l'agent (admin, commercial, webmaster, ...)
• Contexte métier (secteur, contraintes)
• Fuseau horaire client + horaires de travail (obligatoire)
• Site/repo GitHub si l'agent modifie un site

Convention horaire (obligatoire)

Pour toute alerte/rappel client :

• afficher d'abord l'heure client,
• puis conversions utiles (UTC / Taiwan / France) seulement si nécessaire.

Format recommandé : Rappel: 09:00 Europe/Paris (16:00 Taiwan, 08:00 UTC)

Principe d'architecture

• Skill = procédure / playbook
• Plugin = code runtime (nouveaux tools, hooks, commandes)

Pour ce cas d'usage (onboarding client), rester en skill modulaire.