Templates & Agents

Business templates

Six complete astromesh/v1 agents, in templates/. The architect starts from the one closest to your description, and /leia templates <name> shows and explains any of them.

Template	Channel	Pattern	Purpose	Guardrails
`customer-support`	WhatsApp	`react`	FAQ, troubleshooting, complaints, escalation	PII redaction in and out; 1,600 chars
`restaurant-booking`	WhatsApp	`plan_and_execute`	Reservations, menu, dietary needs, cancellations	PII redaction in; 1,600 chars
`ecommerce-assistant`	WhatsApp	`react`	Product search, prices, order tracking, returns; never takes payments	PII redaction in and out; 1,600 chars
`appointment-scheduler`	WhatsApp	`plan_and_execute`	Booking, rescheduling, cancellations	1,600 chars
`lead-qualifier`	WhatsApp	`react`	Qualifies leads with BANT and hands qualified ones to sales	PII redaction in; 1,600 chars
`onboarding-guide`	Web	`react`	New-hire orientation: policies, IT setup, benefits	PII redaction in and out; 4,000 chars

All of them:

use Ollama llama3.1:8b, so they run with no cloud key, and the single-model primary shape that the runtime still accepts;
nest memory as spec.memory.conversational and cap length with max_chars. Older versions used a flat memory block and limit, which the runtime silently ignored.

Change the model before publishing to a hub that does not serve Ollama.

The interpreter matches your words to a template: restaurant, menu or table go to restaurant-booking; shop, product or order to ecommerce-assistant; support, ticket or complaint to customer-support; appointment, clinic or salon to appointment-scheduler; sales, leads or CRM to lead-qualifier; HR, new hire or onboarding to onboarding-guide. Anything else is custom.

The subagents

Subagent	Model	Tools	Role
`leia-interpreter`	Sonnet	Read, Glob	Classifies the request and extracts entities; never invents a name or tenant
`leia-architect`	Opus	Read, Glob, Bash, Write	Designs the manifest and writes it
`leia-operator`	Sonnet	Read, Bash, Write, Glob	Talks to Nexus (`curl`) and the cluster (`kubectl`)
`leia-tester`	Sonnet	Read, Bash, Glob	Chats with the agent or runs scored scenarios
`leia-doctor`	Sonnet	Read, Bash, Glob	Diagnoses with a fixed checklist

What the architect will and won’t write

It will:

use the six real patterns: react, plan_and_execute, pipeline, parallel_fan_out, supervisor, swarm;
use the real sections: spec.prompts.system (not spec.system_prompt) and spec.guardrails (not spec.constraints);
offer a client tool when the value is the call itself (show a chart, open a form) and nothing needs to run on the server;
use per-role models only with a cloud key and a pattern with separate roles, always with default, and always with api_key_env on litellm candidates;
offer spec.output_schema when something downstream reads a value, and spec.chain when you describe what happens after the agent finishes. It states that both need core 0.38.1+ and whether it checked the runtime’s version.

It won’t:

emit internal, webhook, rag or mcp_* tool types, which the runtime drops with only a warning;
invent a spec.channels block;
write a placeholder system prompt;
add a chain because it looks sophisticated.

The bundled schemas

The architect reads these before designing, and they are useful to read directly:

File	Covers
`schemas/astromesh-v1-agent.md`	The agent manifest: identity, per-role models and which keys each source actually consumes, prompts, patterns, tools that load, memory, guardrails, `output_schema` with its supported and ignored keywords, `chain`, the streaming contract and `usage.by_model`
`schemas/orchestration-patterns.md`	When to use each pattern, the roles each one asks for, and why a pattern is not a chain
`schemas/whatsapp-config.md`	WhatsApp variables, the per-agent webhook path, `contact_name`, delivery receipts
`schemas/nexus-api.md`	The Nexus 0.3 API: dual auth, users, tenants, keys, agents, and the runtime’s RAG and workflow endpoints