Imported from chris-dare-dev/agent-kit (data/claude-md/AGENTS.md). Install upstream with npx skills add chris-dare-dev/agent-kit --skill claude-md. Copyright stays with the author.

AGENTS.md — agent registry (generated router)

The agent registry, generated as a thin provider-neutral router: the full agent/skill/entrypoint inventory is MCP-served (below); only normative dispatch, coordination and path rules stay inline.

Full inventory via MCP (not inlined)

• Agents (20) — list_agents / get_agent({name})
• Skills (1) — list_skills / get_skill({name})
• Entrypoints (5) — slash commands in .claude/commands/
• Agent families — 20 agents across 5 domains; enumerate/filter via list_agents
• Canonical machine catalog — data/facts/catalog.json (catalog-generate.py)
• Provider degradation — see CONTEXT.md (Codex/OpenCode-partial items)

Dispatch patterns

• Orchestrator + Workers — 1 orchestrator does deep config analysis (Helm values, overlays, rendered manifests) while 3-6 workers run parallel kubectl diagnostics across clusters/namespaces (ArgoCD sync, Envoy xDS, pod logs, SE/DR state).
• Parallel Cluster Triage — for an issue spanning several clusters or environments, dispatch one diagnostic agent per cluster simultaneously, not sequentially.
• Post-deployment verification — after new ServiceEntries/DestinationRules: confirm ArgoCD Synced+Healthy, confirm the LIVE spec matches the rendered manifest (not just sync status), and rollout restart pods predating the SE so they pick up new Envoy clusters.
• Non-overlapping parallel fix — when one symptom has independent root causes across repos, dispatch agents with DISJOINT repo ownership: name each agent's owned paths + its must-not-touch set; pre-state what's already diagnosed; grant push up-front; require a concise structured report (commit SHA/URL + change + verification). Split within one repo runs sequentially — same-repo commits race the index.

Cross-agent coordination

• Specialist agents that carry only Read/Glob/Grep/Bash cannot call MCP tools — check the agent's declared tool set before routing work that needs one. MCP reads and writes happen in the MAIN session or via skills.
• After editing a platform repo, push routine commits to main after authorization; use no feature branch/MR unless Chris asks. dev is retired. CI renders to deploy:dev; Kargo promotes deploy:stage+deploy:main.
• Agent-artifact finalization — use canonical skills; after terminal validation, refresh Obsidian and emit the append-only receipt. Qdrant is derived search; Graphiti stays candidate-only/write-disabled pending explicit approval. Preserve sources and report capture failures; never write sinks.

Path & subdirectory references

• Agent definitions (canonical): data/agents/*.md (planted at .claude/agents/; enumerate ls .claude/agents/), or list_agents / get_agent({name}).
• Skills: .claude/skills/ (enumerate ls .claude/skills/), or list_skills / get_skill({name}).