agent-runtime-packaging

name	agent-runtime-packaging
description	Build and maintain product-grade agent infrastructure: onboarding diagnostics, runtime home, packaging, skill lifecycle, tool manifests, and agent scaffolds.
triggers	["agent runtime packaging","use agent runtime packaging","run agent runtime packaging"]

Agent Runtime Packaging

Use this skill when turning Bravo, Atlas, Maven, Aura, Hermes, or a client agent from a repo into an installable, diagnosable, repeatable product.

The goal is Hermes-grade infrastructure under Bravo-grade business intelligence.

Core Principle

Business intelligence sits on top of infrastructure. The infrastructure should make the agent easy to install, inspect, operate, repair, extend, and package for another business.

Required Layers

Every productized agent should have:

1. Entry point docs: AGENTS.md, CLAUDE.md, GEMINI.md, ANTIGRAVITY.md.
2. Brain and memory: brain/, memory/, state, active tasks, session log.
3. CLI layer: setup, doctor, status, tools, skills, browser, sessions, gateway.
4. Runtime home: ~/.bravo, ~/.atlas, ~/.maven, ~/.aura, or client-specific equivalent.
5. Config template: no secrets, clear env requirements, safe defaults.
6. Skill lifecycle: create, register, validate, audit, promote, deprecate.
7. Tool manifest: discoverable scripts and commands.
8. Browser intelligence: Browser Harness plus domain skills and approval gates.
9. Safety gates: outbound chokepoint, finance approval, production approval, destructive-action approval.
10. Diagnostics: read-only doctor command that explains what is missing.

Current Bravo Files

• Onboarding doctor: scripts/onboarding_diagnostics.py
• Browser doctor: scripts/browser_harness_doctor.py
• Runtime notes: runtime/README.md
• Install notes: install/README.md
• Config example: config/bravo-config.example.toml
• Browser skill: skills/browser-harness/SKILL.md
• Browser domain memory: browser/domain-skills/

Build Order

When packaging a new agent:

1. Create the brain/memory/entry-point skeleton.
2. Add config template and env scaffold.
3. Add a doctor command before adding automation.
4. Add tool manifest and skill registry.
5. Add Browser Harness domain-skill folder.
6. Add install scripts.
7. Add session store/search.
8. Add gateway adapters only after the basic doctor is green.

Skill Lifecycle

Skills should move through these states:

• draft: local only, not trusted yet.
• registered: discoverable in docs and registry.
• validated: used successfully and passes structural checks.
• promoted: included in the default agent pack.
• deprecated: replaced or unsafe, retained only for history.

Every skill needs:

• YAML frontmatter with name and description.
• A clear trigger.
• Safety constraints.
• Verification steps.
• References to scripts, docs, or templates it uses.

Diagnostics Standard

Doctor commands must be read-only and should print:

• required tools
• missing credentials without revealing values
• required directories/files
• service reachability if safe
• current attach/runtime state
• exact next action

Doctor commands should not:

• mutate .env.agents
• run database migrations
• send messages
• start paid services
• delete or overwrite files
• hide partial failure behind a green status

Safety

Never let packaging work bypass Bravo's business safety:

• Outbound stays behind scripts/send_gateway.py.
• Browser writes require approval.
• Finance actions require approval.
• Production setting changes require approval.
• Destructive actions require approval.

Related

• [[skills/INDEX]]
• [[brain/CAPABILITIES]]