create-collection

name	create-collection
description	Author or refresh `<pack>/.catalog/collection.yaml` and related `.catalog/` artifacts from golden sources (SKILL.md, README, AGENTS.md, Lola marketplace). Use when: - Adding a new pack or refreshing the collection catalog for GitHub Pages / tooling - Aligning catalog narrative, sample workflows, and decision guide with skills on disk - Preparing a PR after changing skills or marketplace metadata Outputs only under `<pack>/.catalog/` (never overwrite README, SKILL, CLAUDE, or marketplace YAML).
model	inherit
color	blue
allowed-tools	Read Glob Grep Bash

Create collection catalog

Audience: Maintainers and assistants updating per-pack .catalog/ data.

Goal: Produce a coherent collection.yaml that passes make validate-collection-compliance, using human judgment for synthesis—not a blind dump of frontmatter.

Prerequisites

• Repository root as cwd.
• Read COLLECTION_SPEC.md and catalog/schema.yaml.
• Optional: run uv run python scripts/scaffold_catalog.py <pack> for a stdout draft roster.

When to Use

• New marketplace module or new pack directory needs .catalog/.
• Skills were added/removed/renamed and roster parity must match disk.
• Marketplace description / version changed and catalog should reflect it (still edit marketplace in git separately).

Workflow

1. Resolve pack — directory name must appear in union(marketplace/modules[].path, docs/plugins.json keys) and exist on disk.

2. Read sources in order (precedence):

   • skills/*/SKILL.md (frontmatter + body for summaries and orchestration hints)
   • <pack>/README.md
   • <pack>/AGENTS.md (intent routing → skills_decision_guide ideas)
   • Matching row in marketplace/rh-agentic-collection.yml (path == pack)

3. Classify skills — place each skill in contents.skills or contents.orchestration_skills using maintainer judgment. Optional hint: metadata.collection.role: orchestration in SKILL.md frontmatter. Names in YAML must match the skills/<name>/ directory name.

4. Write <pack>/.catalog/collection.yaml — start with the standard # banner (see COLLECTION_SPEC). Include every field required by catalog/schema.yaml. Keep inline strings under 500 Unicode code points for monitored fields (summary, documentation_section, mcp_section, security_model, and inline deploy_and_use); otherwise move prose to a sibling .md and set the same field to a one-line ref, e.g. documentation_section: '#documentation_section.md' or deploy_and_use: '#deploy_and_use.md' (see COLLECTION_SPEC). Do not use parallel documentation_section_file / mcp_section_file / security_model_file keys.

   deploy_and_use / install prose: Follow COLLECTION_SPEC.md deploy_and_use content (install + env + MCP). If mcps.json defines MCP servers, document prerequisites, export lines for every ${VAR} name used there, Lola + marketplace path:, optional Claude/Cursor install when those marketplaces are listed, and an MCP configuration subsection. Prefer deploy_and_use: '#deploy_and_use.md' plus deploy_and_use.md structured like rh-sre/.catalog/deploy_and_use.md (see rh-virt/.catalog/deploy_and_use.md for a kubeconfig-focused variant).

   YAML multiline style (| literal blocks): For contents.description, every summary_markdown (regular and orchestration skills), and every sample_workflows[].workflow, use a literal block scalar (field: |) instead of a single-quoted string that continues across many YAML lines with spacer blanks. Indent each body line one level deeper than the key line. Inside the block, keep consecutive lines without empty lines between the intro, **Use when:**, list items, and **What it does:** (and similar headings) unless you intentionally want extra paragraph breaks in rendered markdown—reference rh-sre and rh-developer .catalog/collection.yaml for the house style. Prose with apostrophes (e.g. What's) stays readable in | without '' escaping.

5. Mirror JSON — from repo root: uv run python scripts/catalog_yaml_to_json.py --pack <pack> (or make catalog-mirror-json).

6. Self-review checklist

   • Every on-disk skills/<n>/SKILL.md appears exactly once in skills ∪ orchestration_skills (orchestration-only skills such as /remediation still live under skills/<name>/; list them only under contents.orchestration_skills when they orchestrate others).
   • No TODO: / TBD in sample_workflows.workflow; each workflow includes User: and - bullets.
   • skills_decision_guide empty if the pack has no skills; otherwise each skill_to_use matches a skill dir.
   • resources[].url set; embedded_doc only if that path exists under the pack.
   • Install / deploy (deploy_and_use): inline (≤ 500 code points) or one-line #deploy_and_use.md ref (see COLLECTION_SPEC.md install + env + MCP). If mcps.json has MCP servers: fragment (or inline) must include prerequisites, export examples for each ${VAR} name from mcps.json, Lola/marketplace path:, and MCP configuration notes; add Claude Code / Cursor install when marketplaces lists them. Fragments use the HTML provenance banner. Use mcps.json, not .mcp.json.
   • Other long blocks: documentation_section, mcp_section, and security_model follow deploy_and_use: inline markdown (≤ 500 code points) or one-line #fragment.md on that same key (sibling under .catalog/; no .catalog/ prefix in the YAML string).
   • Publication-style metadata: maturity (required) — GREEN (listed on GitHub Pages catalog) or ORANGE (in-repo only until promoted). Also when useful: support_level, author, homepage, keywords, legal_resources (URLs only).
   • version / listing fields: Align version and core listing copy with the matching marketplace/rh-agentic-collection.yml row (path == pack); do not bump marketplace YAML from the catalog workflow.
   • Multiline fields: contents.description, all summary_markdown, and all sample_workflows[].workflow use | block scalars (not long single-quoted multiline); content lines are compact unless extra blank lines are deliberate for markdown.

7. Validate — make validate-collection-compliance before commit.

Dependencies

• make validate-collection-compliance
• uv run python scripts/scaffold_catalog.py <pack> (optional)

Common Issues

• CI: collection.json drift — regenerate with make catalog-mirror-json.
• Roster errors — YAML name must equal the skill folder name, not a display alias.
• Empty support pack — if there are no skills/, use empty skills, orchestration_skills, and skills_decision_guide: [].
• Unreadable multiline YAML — replace spanning '…' blocks with field: | and indented lines; see step 4 YAML multiline style.
• Thin deploy_and_use — pack has mcps.json but catalog only says “see README”; add Environment setup (export for each ${VAR}) and MCP configuration per COLLECTION_SPEC (use #deploy_and_use.md like rh-sre / rh-virt).

Example usage

uv run python scripts/scaffold_catalog.py rh-sre
uv run python scripts/catalog_yaml_to_json.py --pack rh-sre
make validate-collection-compliance