…ry-dev#229)

# Description
agentregistry supports agents, skills, and MCP servers as first-class
resources, but has no way to manage reusable prompt templates (system
prompts, instructions) that agents can consume. Users currently hardcode
instructions in agent code, with no versioning, sharing, or central
management.

What changed: Adds prompts as a fourth first-class resource type across
all layers of the registry.

**Registry CRUD**

- PromptJSON model — a prompt is a named, versioned text string (name,
description, version, content)
- Database migration (004_add_prompts_table.sql) with the same schema
pattern as skills/agents
- PostgreSQL implementation for all prompt CRUD operations
- Service layer with version management, duplicate prevention, latest
tracking
- REST API endpoints: `GET/POST /v0/prompts`, `GET/DELETE
/v0/prompts/{name}/versions/{version}`
- API client methods for all operations
- Auth resource type PermissionArtifactTypePrompt


**CLI commands**

- `arctl prompt publish <file> --name <name> --version <version>` —
publish from a plain text file
- `arctl prompt list / arctl prompt show <name> / arctl prompt delete
<name> --version <version>`
- `arctl agent add-prompt <local-name> --registry-prompt-name <name>` —
adds a prompt reference to agent.yaml, defaults registryURL to the
current registry

**Agent runtime integration**

- PromptRef type in AgentManifest with registryURL, registryPromptName,
registryPromptVersion
- `arctl agent run` resolves prompt refs from the registry REST API,
writes prompts.json to the agent config directory
- `prompts_loader.py` (autogenerated, refreshed on every run) loads
prompts and exposes `build_instruction(default)` — returns registry
prompt content if available, falls back to the default
- Agent template uses `instruction=build_instruction("default
instruction here")`

**MCP server**

- Adds user-facing MCP prompts (search_registry, deploy_mcp_server,
registry_overview) that guide users through the registry's own tools —
following the MCP spec's intent for prompts as user-controlled slash
commands
- Registry prompts are NOT forwarded via the MCP prompts protocol
(they're agent-consumed data, not user-facing slash commands)


# Change Type
```
/kind feature
```

# Changelog

```release-note
Add prompts as a first-class registry resource with full CRUD support (REST API, CLI, database), agent manifest integration (`arctl agent add-prompt`), and runtime resolution during `arctl agent run`. Prompts are versioned text templates that agents use as instructions, resolved from the registry at launch time.
```

**Additional Notes**

- Existing agents scaffolded before this change need a one-line update
to `agent.py`: replace the hardcoded instruction="..." with
instruction=build_instruction("...") and add `from .prompts_loader
import build_instruction`. New agents get this automatically from the
updated template.

- The database migration (`004`) creates the `prompts` table. Existing
databases will auto-migrate on server restart.
- Prompt content is stored as a plain string in the registry's JSONB
value column. The PromptJSON model intentionally avoids MCP-specific
fields (arguments, messages) since prompts are consumed as agent
instructions, not as MCP protocol prompts.

---------

Signed-off-by: Peter Jausovec <[email protected]>
Co-authored-by: Claude Opus 4.6 (1M context) <[email protected]>